Nabble - jikesrvm-researchers

Re: [rvm-research] Using PIN in Jikes RVM 3.0.0

2009-11-23T13:54:31Z

Hi Raj,

It looks like you're trying to load a 64-bit pintool with a 32-bit version of pin, which won't work.

When I did this (about a year ago now), I had the best luck running under a 32-bit OS -- pin really got confused in the presence of multilib. I also had a much easier time using native threads instead of m:n threads (native threads were only available as a prerelease at the time, but they are the default in 3.1.0.)

Since I was running dacapo under Pin, I also used a simple callback class that made bogus system calls before and after benchmark runs (as an out-of-band way to communicate with my pintool to start and stop collecting data). The other thing you'll have to worry about is that Jikes maps out a large part of the virtual address space, so your pintool won't be able to allocate that much data. If you need to deal with a lot of data, consider having your pintool pipe events to another application that can process them and keep them in memory.

Some details about what I did with Jikes and Pin are in chapter 5 of my dissertation, which is available here: http://web.willbenton.com/research/dissertation

best,
wb

On Nov 19, 2009, at 11:05 PM, Rajkishore Barik wrote:

> Hi,
>
> I am trying to use PIN in Jikes RVM 3.0.0 (32bit application) on a 64-
> bit Redhat-Linux OS.
> I am a newbie for PIN. Can someone tell which version of PIN should I
> download and what
> command should I use to instrument in jikesrvm ? I downloaded the 64-
> bit and used the following
> command:
>
> rvm -wrap "$HOME/tools/pin-2.6/ia32/bin/pinbin -t $HOME/tools/pin-2.6/
> source/tools/SimpleExamples/obj-intel64/icount.so -- " Filename
>
> I get the following error:
> E:Unable to load $HOME/tools/pin-2.6/source/tools/SimpleExamples/obj-
> intel64/icount.so
>
> Best Regards,
> Raj
>
>
> ------------------------------------------------------------------------------
> Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
> trial. Simplify your report design, integration and deployment - and focus on
> what you do best, core application coding. Discover what's new with
> Crystal Reports now. http://p.sf.net/sfu/bobj-july
> _______________________________________________
> Jikesrvm-researchers mailing list
> Jikesrvm-researchers@...
> https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] where object allocation on heap requested by a thread?

2009-11-23T10:56:17Z

Hi,

An allocation in Jikes RVM looks like this for a scalar value (not an
array). First, the compiler inserts a call to resolvedNewScalar in
org.jikesrvm.runtime.RuntimeEntrypoints in the compiled code. Then, the
call chain looks like this:
resolvedNewScalar in org.jikesrvm.runtime.RuntimeEntrypoints ->
allocateScalar in org.jikesrvm.mm.mminterface.MemoryManager ->
allocateSpace in org.jikesrvm.mm.mminterface.MemoryManager ->
alloc in org.jikesrvm.mm.mminterface.Selected.Mutator

Selected.Mutator derives from whatever MMTk collector has been selected
in the compile-time configuration. So for example in the production
configuration, it will call the alloc method in
org.mmtk.plan.generational.immix.GenImmixMutator.

-Eddie

pankajsejwal wrote:
> I found out that with the help of eclipse it is not possible to find
> references to allocation method of any mutator in any GC because it seems
> like rvm and mmtk are two seperate folders.
> I am trying to find the method where any thread tries to call mmtk(though it
> has not been imported,only mm interface is there)to allocate an object
> greater than its stack size on heap.
> please guide me..

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] Detailed warnings

2009-11-23T07:08:30Z

Luke Terry <ljt06@...> wrote on 11/20/2009 05:15:32 AM: > > I've been doing modifications to the JikesRVM and after the build it > reports the number of warnings, but I cannot find the details of the > warnings anywhere. > Where/how do I view the warnings themselves? >
If the warnings are coming from the bootimage writing process, you'll be able to find them in a file called BootImageWriterOutput.txt that should be sitting the same directory as the rvm image that you built.

--dave

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] About the size of region and poll frequency

2009-11-23T00:14:31Z

lltong wrote:

> Thanks very much.
>
> I remember you mentioned MMTk test harness before. Actually just now I
> was trying to play with it, unfortunately it fails. In
> org.mmtk.harness.Main, there import two classes, which cannot be found
> in the source code of RVM:
>
> import org.mmtk.harness.lang.parser.Parser;
> import org.mmtk.harness.lang.parser.ParseException;
>
> Am I wrong at any places?

Figured it out already, thanks.

>
> Thanks.
>
>
> Robin Garner wrote:
>> lltong wrote:
>>
>>> en, thanks very much.
>>>
>>> One more inquiry about the relation between poll() and acquire().
>>> Take MarkCompact as an example, if I remove boolean heapFull in
>>> poll(), it just won't collect. Yet it will work happily and
>>> correctly if boolean mustCollect are removed. Could you explain why?
>>>
>>
>> Which benchmark, which heap size ?
>>
>> Obviously in the configuration you are testing, collection is
>> generally triggered by running up against a 'heap full' condition.
>> If you find a benchmark that allocates lots of large objects, and run
>> it in a very large heap, mustCollect will become important.
>>
>>> In my understanding, the headers of regions remain unchanged even
>>> after GC. It means that, ever since the first collection, the whole
>>> MC space is segregated into various regions allocationed in the
>>> first mutation. These regions will not be zeroed and continue to be
>>> reused in the following mutations.
>>>
>>
>> Maybe Daniel could answer this - I'm not that familiar with MC.
>>
>>
>>> In BumpPointer, every slow-path allocation will first check whether
>>> the nextRegion is null. If not, Allocation won't call
>>> Space.acquire() to draw a new region but reuse the old nextRegion.
>>> Because region headers won't be cleared once created, the nextRegion
>>> is always not null.
>>
>> Not necessarily true. In the SemiSpace collector, for example, each
>> of the spaces is discarded after each collection.
>>
>> You seem to be reasoning from the specific to the general - keep in
>> mind there are many collectors in MMTk, and it's designed as a
>> toolkit, ie with collectors in mind that may not yet exist.
>>
> you are right.
>>
>>> Therefore, in the following mutations, Space.acquire() won't be
>>> called. But Space.acquire() is the only function that call poll() to
>>> evaluate the necessity of a GC. How GC can be triggered in this
>>> situation?
>>>
>>
>> MC is not the most stable of collectors. I'd ask you to reproduce it
>> in 3.1.0, but MC doesn't work there. Is MarkCompact important to you
>> ? Why not use SemiSpace or MarkSweep ?
>>
> yeah..I am working on improving MC algorithms.
> I think I figured out the reason here. Actually after the whole space
> is segregated into regions, calling to acquire() comes to be scarce.
> Actually only two each mutation: the first allocation and the last
> allocation of this mutation. Like the following:
>
> [GC 8 Start 12.90 s 25632KB -> 8000 KB 931.59 ms]
> acquire is called
> 1 9 7 4 8 5
> + 5 2 6 4 8 5
> ------
> = 7 2 3 9 7 0
> check #9= 22
> acquire is called
> [GC 9 Start 14.52 s 25632KB -> 7936 KB 926.35 ms]
> acquire is called
> 1 9 7 4 8 5
> + 5 2 6 4 8 5
> ------
> = 7 2 3 9 7 0
> check #10= 23
> acquire is called
>
> Upon the first allocation, initialRegion is reset to be zero thus
> require acquire() function to fetch the first region(which is created
> by previous mutations); While the NEXT_REGION_OFFSET of the last
> region is zero, therefore calls another acquire(), which results in a
> space full garbage collection.
>>
>>> I tried to add Log.writeln("acquire is called") after
>>> Space.acquire() function call in BumpPointer.allocSlowOnce(), to
>>> record the usage of acquire() function. But this message didn't
>>> display while running benchmarks. Could you tell me how to achieve
>>> my goal?
>>>
>>
>> Which collector? which benchmark? what heap size ?
>>
> Got the reason, I am using my revised version of BumpPointer.java but
> compiling the MC collectors with the original version : (
>

--
Regards,
Liangliang

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] About the size of region and poll frequency

2009-11-22T23:51:25Z

Thanks very much.

I remember you mentioned MMTk test harness before. Actually just now I
was trying to play with it, unfortunately it fails. In
org.mmtk.harness.Main, there import two classes, which cannot be found
in the source code of RVM:

import org.mmtk.harness.lang.parser.Parser;
import org.mmtk.harness.lang.parser.ParseException;

Am I wrong at any places?

Thanks.

Robin Garner wrote:

> lltong wrote:
>
>> en, thanks very much.
>>
>> One more inquiry about the relation between poll() and acquire(). Take
>> MarkCompact as an example, if I remove boolean heapFull in poll(), it
>> just won't collect. Yet it will work happily and correctly if boolean
>> mustCollect are removed. Could you explain why?
>>
>
> Which benchmark, which heap size ?
>
> Obviously in the configuration you are testing, collection is generally
> triggered by running up against a 'heap full' condition. If you find a
> benchmark that allocates lots of large objects, and run it in a very
> large heap, mustCollect will become important.
>
>> In my understanding, the headers of regions remain unchanged even after
>> GC. It means that, ever since the first collection, the whole MC space
>> is segregated into various regions allocationed in the first mutation.
>> These regions will not be zeroed and continue to be reused in the
>> following mutations.
>>
>
> Maybe Daniel could answer this - I'm not that familiar with MC.
>
>
>> In BumpPointer, every slow-path allocation will first check whether the
>> nextRegion is null. If not, Allocation won't call Space.acquire() to
>> draw a new region but reuse the old nextRegion. Because region headers
>> won't be cleared once created, the nextRegion is always not null.
>>
>
> Not necessarily true. In the SemiSpace collector, for example, each of
> the spaces is discarded after each collection.
>
> You seem to be reasoning from the specific to the general - keep in mind
> there are many collectors in MMTk, and it's designed as a toolkit, ie
> with collectors in mind that may not yet exist.
>

you are right.

>
>> Therefore, in the following mutations, Space.acquire() won't be called.
>> But Space.acquire() is the only function that call poll() to evaluate
>> the necessity of a GC. How GC can be triggered in this situation?
>>
>
> MC is not the most stable of collectors. I'd ask you to reproduce it in
> 3.1.0, but MC doesn't work there. Is MarkCompact important to you ?
> Why not use SemiSpace or MarkSweep ?
>

yeah..I am working on improving MC algorithms.
I think I figured out the reason here. Actually after the whole space is
segregated into regions, calling to acquire() comes to be scarce.
Actually only two each mutation: the first allocation and the last
allocation of this mutation. Like the following:

[GC 8 Start 12.90 s 25632KB -> 8000 KB 931.59 ms]
acquire is called
1 9 7 4 8 5
+ 5 2 6 4 8 5
------
= 7 2 3 9 7 0
check #9= 22
acquire is called
[GC 9 Start 14.52 s 25632KB -> 7936 KB 926.35 ms]
acquire is called
1 9 7 4 8 5
+ 5 2 6 4 8 5
------
= 7 2 3 9 7 0
check #10= 23
acquire is called

Upon the first allocation, initialRegion is reset to be zero thus
require acquire() function to fetch the first region(which is created by
previous mutations); While the NEXT_REGION_OFFSET of the last region is
zero, therefore calls another acquire(), which results in a space full
garbage collection.
>
>> I tried to add Log.writeln("acquire is called") after Space.acquire()
>> function call in BumpPointer.allocSlowOnce(), to record the usage of
>> acquire() function. But this message didn't display while running
>> benchmarks. Could you tell me how to achieve my goal?
>>
>
> Which collector? which benchmark? what heap size ?
>
Got the reason, I am using my revised version of BumpPointer.java but
compiling the MC collectors with the original version : (

--
Regards,
Liangliang

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] About the size of region and poll frequency

2009-11-22T22:59:32Z

lltong wrote:
> en, thanks very much.
>
> One more inquiry about the relation between poll() and acquire(). Take
> MarkCompact as an example, if I remove boolean heapFull in poll(), it
> just won't collect. Yet it will work happily and correctly if boolean
> mustCollect are removed. Could you explain why?

Which benchmark, which heap size ?

Obviously in the configuration you are testing, collection is generally
triggered by running up against a 'heap full' condition. If you find a
benchmark that allocates lots of large objects, and run it in a very
large heap, mustCollect will become important.

> In my understanding, the headers of regions remain unchanged even after
> GC. It means that, ever since the first collection, the whole MC space
> is segregated into various regions allocationed in the first mutation.
> These regions will not be zeroed and continue to be reused in the
> following mutations.

Maybe Daniel could answer this - I'm not that familiar with MC.

> In BumpPointer, every slow-path allocation will first check whether the
> nextRegion is null. If not, Allocation won't call Space.acquire() to
> draw a new region but reuse the old nextRegion. Because region headers
> won't be cleared once created, the nextRegion is always not null.

Not necessarily true. In the SemiSpace collector, for example, each of
the spaces is discarded after each collection.

You seem to be reasoning from the specific to the general - keep in mind
there are many collectors in MMTk, and it's designed as a toolkit, ie
with collectors in mind that may not yet exist.

> Therefore, in the following mutations, Space.acquire() won't be called.
> But Space.acquire() is the only function that call poll() to evaluate
> the necessity of a GC. How GC can be triggered in this situation?

MC is not the most stable of collectors. I'd ask you to reproduce it in
3.1.0, but MC doesn't work there. Is MarkCompact important to you ?
Why not use SemiSpace or MarkSweep ?

> I tried to add Log.writeln("acquire is called") after Space.acquire()
> function call in BumpPointer.allocSlowOnce(), to record the usage of
> acquire() function. But this message didn't display while running
> benchmarks. Could you tell me how to achieve my goal?

Which collector? which benchmark? what heap size ?

> Thanks.
>
> Best Regards,
> Liangliang
>
> Robin Garner wrote:
>> lltong wrote:
>>
>>> Dear all,
>>>
>>> 1) As I can understand, region is allocated by /chunkSize/ in
>>> /BumpPointer.java/, detailed in the following:
>>>
>>> /Extent chunkSize = Word.fromIntZeroExtend(bytes).plus(CHUNK_MASK)
>>> .and(CHUNK_MASK.not()).toExtent();/
>>>
>> This expression rounds a request up to the next highest multiple of the
>> chunk size.
>>
>>
>>> if the current requested bytes are less than /CHUNK_MASK/ (32K - 1), the
>>> bytes will be masked away. Yet objects larger than 8KB is allocated
>>> into LOS space, does that mean /MonotonePageResource/ is cut equally
>>> into 32K-size regions while satisfying allocations?
>>>
>> The code in BumpPointer makes no assumptions about how the Plan as a
>> whole is structured. If you wanted to, you could construct a collector
>> using a single semispace, with no LOS.
>>
>> The code in BumpPointer is concerned with implementing an allocator that
>> can satisfy any allocation request (it's a *mechanism*). The decision
>> to allocate large objects into the LOS is a *policy* decision,
>> independent of how BumpPointer is implemented.
>>
>
> I see, a good design framework it is.
>> A MonotonePageResource allocates pages monotonically, in the manner they
>> are requested. There are several spaces that use MonotonePageResource,
>> and each of them consumes pages differently. In the case of a
>> BumpPointer, pages are consumed in multiples of the BLOCK_MASK.
>>
>>
>>> 2) in /PageResource.java/ there is an /reservePages(int pages)/ function
>>> like this:
>>>
>>> /public final boolean reservePages(int pages) throws InlinePragma {
>>> lock();
>>> reserved = committed + pages;
>>> boolean satisfied = reserved <= pageBudget;
>>> unlock();
>>> return satisfied;
>>> }/
>>>
>>> But the value of /pageBudget/ is passed from /DEFAULT_POLL_FREQUENCY/,
>>> which equals to 32 pages. As I can see, it means that ever since
>>> /reserved/ exceeds 32 pages, this function will always return /false /to
>>> function /acquire(int pages)/ in /Space.java/. It means upon every
>>> allocation the /poll/ function will be called. Why should we poll so
>>> frequently?
>>>
>> At first glance, there looks to be a bug in this code - even if pages
>> are allocated in smaller chunks, we will poll for every allocation
>> beyond the first 'pageBudget' pages.
>>
>> Allocating every 32 pages of allocation isn't particularly frequent,
>> given the cost of the poll function. Remember this is called from the
>> slowpath, which is not particularly performance critical.
>>
> en..that makes sense.
>>
>>> 3) Still in /poll/ function, it says collection will be triggered if
>>> mustCollect or heapFull. /mustCollect /is true when /Space.acquire()/
>>> returns false, which means the allocation cursor reaches the sentinel of
>>> the /PageResource/ associated with this /Space/. In my understanding, if
>>> /mustCollect/ then heap must be full. I tried to remove heapFull, then
>>> GC starts works awkwardly. Can you tell me why heapFull is a must?
>>>
>> The poll() function has been refactored somewhat over time, and I think
>> it's easier to understand in the svn HEAD.
>>
> I see, unfortunately I need some codes which are written for Jikesrvm
> long long ago, so I have to resort to old versions.
>> There are several conditions under which a GC can be required. One is
>> that a given Space is full (which happened a lot before the introduction
>> of discontiguous spaces). If 100MB (for example) is allocated to the
>> LOS, then despite the heap size settings, when 100MB of large objects
>> have been allocated, a collection must occur, no matter how much heap is
>> available in the other Spaces. When a space is full, poll() is invoked
>> with mustCollect=true.
>>
>> On the other hand, if the heap size is set to a given size, then we must
>> collect when the cumulative size of all spaces reaches that total (or
>> before, to allow for copy reserve and other factors).
>>
> Thanks so much.
>> hope this helps,
>> Robin
>>
>>
>>> Thanks so much.
>>>
>>>
>>>
>>
>> ------------------------------------------------------------------------------
>> Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
>> trial. Simplify your report design, integration and deployment - and focus on
>> what you do best, core application coding. Discover what's new with
>> Crystal Reports now. http://p.sf.net/sfu/bobj-july
>> _______________________________________________
>> Jikesrvm-researchers mailing list
>> Jikesrvm-researchers@...
>> https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers
>>
>
>

Re: [rvm-research] About the size of region and poll frequency

2009-11-22T20:03:24Z

en, thanks very much.

One more inquiry about the relation between poll() and acquire(). Take
MarkCompact as an example, if I remove boolean heapFull in poll(), it
just won't collect. Yet it will work happily and correctly if boolean
mustCollect are removed. Could you explain why?

In my understanding, the headers of regions remain unchanged even after
GC. It means that, ever since the first collection, the whole MC space
is segregated into various regions allocationed in the first mutation.
These regions will not be zeroed and continue to be reused in the
following mutations.

In BumpPointer, every slow-path allocation will first check whether the
nextRegion is null. If not, Allocation won't call Space.acquire() to
draw a new region but reuse the old nextRegion. Because region headers
won't be cleared once created, the nextRegion is always not null.
Therefore, in the following mutations, Space.acquire() won't be called.
But Space.acquire() is the only function that call poll() to evaluate
the necessity of a GC. How GC can be triggered in this situation?

I tried to add Log.writeln("acquire is called") after Space.acquire()
function call in BumpPointer.allocSlowOnce(), to record the usage of
acquire() function. But this message didn't display while running
benchmarks. Could you tell me how to achieve my goal?

Thanks.

Best Regards,
Liangliang

Robin Garner wrote:

> lltong wrote:
>
>> Dear all,
>>
>> 1) As I can understand, region is allocated by /chunkSize/ in
>> /BumpPointer.java/, detailed in the following:
>>
>> /Extent chunkSize = Word.fromIntZeroExtend(bytes).plus(CHUNK_MASK)
>> .and(CHUNK_MASK.not()).toExtent();/
>>
>
> This expression rounds a request up to the next highest multiple of the
> chunk size.
>
>
>> if the current requested bytes are less than /CHUNK_MASK/ (32K - 1), the
>> bytes will be masked away. Yet objects larger than 8KB is allocated
>> into LOS space, does that mean /MonotonePageResource/ is cut equally
>> into 32K-size regions while satisfying allocations?
>>
>
> The code in BumpPointer makes no assumptions about how the Plan as a
> whole is structured. If you wanted to, you could construct a collector
> using a single semispace, with no LOS.
>
> The code in BumpPointer is concerned with implementing an allocator that
> can satisfy any allocation request (it's a *mechanism*). The decision
> to allocate large objects into the LOS is a *policy* decision,
> independent of how BumpPointer is implemented.
>

I see, a good design framework it is.

> A MonotonePageResource allocates pages monotonically, in the manner they
> are requested. There are several spaces that use MonotonePageResource,
> and each of them consumes pages differently. In the case of a
> BumpPointer, pages are consumed in multiples of the BLOCK_MASK.
>
>
>> 2) in /PageResource.java/ there is an /reservePages(int pages)/ function
>> like this:
>>
>> /public final boolean reservePages(int pages) throws InlinePragma {
>> lock();
>> reserved = committed + pages;
>> boolean satisfied = reserved <= pageBudget;
>> unlock();
>> return satisfied;
>> }/
>>
>> But the value of /pageBudget/ is passed from /DEFAULT_POLL_FREQUENCY/,
>> which equals to 32 pages. As I can see, it means that ever since
>> /reserved/ exceeds 32 pages, this function will always return /false /to
>> function /acquire(int pages)/ in /Space.java/. It means upon every
>> allocation the /poll/ function will be called. Why should we poll so
>> frequently?
>>
>
> At first glance, there looks to be a bug in this code - even if pages
> are allocated in smaller chunks, we will poll for every allocation
> beyond the first 'pageBudget' pages.
>
> Allocating every 32 pages of allocation isn't particularly frequent,
> given the cost of the poll function. Remember this is called from the
> slowpath, which is not particularly performance critical.
>

en..that makes sense.

>
>> 3) Still in /poll/ function, it says collection will be triggered if
>> mustCollect or heapFull. /mustCollect /is true when /Space.acquire()/
>> returns false, which means the allocation cursor reaches the sentinel of
>> the /PageResource/ associated with this /Space/. In my understanding, if
>> /mustCollect/ then heap must be full. I tried to remove heapFull, then
>> GC starts works awkwardly. Can you tell me why heapFull is a must?
>>
>
> The poll() function has been refactored somewhat over time, and I think
> it's easier to understand in the svn HEAD.
>

I see, unfortunately I need some codes which are written for Jikesrvm
long long ago, so I have to resort to old versions.

> There are several conditions under which a GC can be required. One is
> that a given Space is full (which happened a lot before the introduction
> of discontiguous spaces). If 100MB (for example) is allocated to the
> LOS, then despite the heap size settings, when 100MB of large objects
> have been allocated, a collection must occur, no matter how much heap is
> available in the other Spaces. When a space is full, poll() is invoked
> with mustCollect=true.
>
> On the other hand, if the heap size is set to a given size, then we must
> collect when the cumulative size of all spaces reaches that total (or
> before, to allow for copy reserve and other factors).
>

Thanks so much.

> hope this helps,
> Robin
>
>
>> Thanks so much.
>>
>>
>>
>
>
> ------------------------------------------------------------------------------
> Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
> trial. Simplify your report design, integration and deployment - and focus on
> what you do best, core application coding. Discover what's new with
> Crystal Reports now. http://p.sf.net/sfu/bobj-july
> _______________________________________________
> Jikesrvm-researchers mailing list
> Jikesrvm-researchers@...
> https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers
>

Re: [rvm-research] About the size of region and poll frequency

2009-11-22T18:52:16Z

lltong wrote:
> Dear all,
>
> 1) As I can understand, region is allocated by /chunkSize/ in
> /BumpPointer.java/, detailed in the following:
>
> /Extent chunkSize = Word.fromIntZeroExtend(bytes).plus(CHUNK_MASK)
> .and(CHUNK_MASK.not()).toExtent();/

This expression rounds a request up to the next highest multiple of the
chunk size.

> if the current requested bytes are less than /CHUNK_MASK/ (32K - 1), the
> bytes will be masked away. Yet objects larger than 8KB is allocated
> into LOS space, does that mean /MonotonePageResource/ is cut equally
> into 32K-size regions while satisfying allocations?

The code in BumpPointer makes no assumptions about how the Plan as a
whole is structured. If you wanted to, you could construct a collector
using a single semispace, with no LOS.

The code in BumpPointer is concerned with implementing an allocator that
can satisfy any allocation request (it's a *mechanism*). The decision
to allocate large objects into the LOS is a *policy* decision,
independent of how BumpPointer is implemented.

A MonotonePageResource allocates pages monotonically, in the manner they
are requested. There are several spaces that use MonotonePageResource,
and each of them consumes pages differently. In the case of a
BumpPointer, pages are consumed in multiples of the BLOCK_MASK.

> 2) in /PageResource.java/ there is an /reservePages(int pages)/ function
> like this:
>
> /public final boolean reservePages(int pages) throws InlinePragma {
> lock();
> reserved = committed + pages;
> boolean satisfied = reserved <= pageBudget;
> unlock();
> return satisfied;
> }/
>
> But the value of /pageBudget/ is passed from /DEFAULT_POLL_FREQUENCY/,
> which equals to 32 pages. As I can see, it means that ever since
> /reserved/ exceeds 32 pages, this function will always return /false /to
> function /acquire(int pages)/ in /Space.java/. It means upon every
> allocation the /poll/ function will be called. Why should we poll so
> frequently?

At first glance, there looks to be a bug in this code - even if pages
are allocated in smaller chunks, we will poll for every allocation
beyond the first 'pageBudget' pages.

Allocating every 32 pages of allocation isn't particularly frequent,
given the cost of the poll function. Remember this is called from the
slowpath, which is not particularly performance critical.

> 3) Still in /poll/ function, it says collection will be triggered if
> mustCollect or heapFull. /mustCollect /is true when /Space.acquire()/
> returns false, which means the allocation cursor reaches the sentinel of
> the /PageResource/ associated with this /Space/. In my understanding, if
> /mustCollect/ then heap must be full. I tried to remove heapFull, then
> GC starts works awkwardly. Can you tell me why heapFull is a must?

The poll() function has been refactored somewhat over time, and I think
it's easier to understand in the svn HEAD.

There are several conditions under which a GC can be required. One is
that a given Space is full (which happened a lot before the introduction
of discontiguous spaces). If 100MB (for example) is allocated to the
LOS, then despite the heap size settings, when 100MB of large objects
have been allocated, a collection must occur, no matter how much heap is
available in the other Spaces. When a space is full, poll() is invoked
with mustCollect=true.

On the other hand, if the heap size is set to a given size, then we must
collect when the cumulative size of all spaces reaches that total (or
before, to allow for copy reserve and other factors).

hope this helps,
Robin

> Thanks so much.
>
>

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

where object allocation on heap requested by a thread?

2009-11-22T10:21:58Z

I found out that with the help of eclipse it is not possible to find references to allocation method of any mutator in any GC because it seems like rvm and mmtk are two seperate folders.
I am trying to find the method where any thread tries to call mmtk(though it has not been imported,only mm interface is there)to allocate an object greater than its stack size on heap.I am trying to seek information in RVMThread.java. Is it the right place.
please guide me..

[rvm-research] About the size of region and poll frequency

2009-11-22T06:12:55Z

Dear all,

1) As I can understand, region is allocated by /chunkSize/ in
/BumpPointer.java/, detailed in the following:

/Extent chunkSize = Word.fromIntZeroExtend(bytes).plus(CHUNK_MASK)
.and(CHUNK_MASK.not()).toExtent();/

if the current requested bytes are less than /CHUNK_MASK/ (32K - 1), the
bytes will be masked away. Yet objects larger than 8KB is allocated
into LOS space, does that mean /MonotonePageResource/ is cut equally
into 32K-size regions while satisfying allocations?

2) in /PageResource.java/ there is an /reservePages(int pages)/ function
like this:

/public final boolean reservePages(int pages) throws InlinePragma {
lock();
reserved = committed + pages;
boolean satisfied = reserved <= pageBudget;
unlock();
return satisfied;
}/

But the value of /pageBudget/ is passed from /DEFAULT_POLL_FREQUENCY/,
which equals to 32 pages. As I can see, it means that ever since
/reserved/ exceeds 32 pages, this function will always return /false /to
function /acquire(int pages)/ in /Space.java/. It means upon every
allocation the /poll/ function will be called. Why should we poll so
frequently?

3) Still in /poll/ function, it says collection will be triggered if
mustCollect or heapFull. /mustCollect /is true when /Space.acquire()/
returns false, which means the allocation cursor reaches the sentinel of
the /PageResource/ associated with this /Space/. In my understanding, if
/mustCollect/ then heap must be full. I tried to remove heapFull, then
GC starts works awkwardly. Can you tell me why heapFull is a must?

Thanks so much.

--
Regards,
Liangliang

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] Opt compiler options not working

2009-11-20T07:43:15Z

Thanks a lot,
By using TestHarness im now able to print hir and do other opt-related stuff. Like you had guessed, the test case was not big enough.

On Fri, Nov 20, 2009 at 10:45 AM, David P Grove <groved@...> wrote:

If you are trying to write small test cases to see what the opt compiler is doing, you might want to learn about OptTestHarness (see the userguide, online at http://jikesrvm.org/OptTestHarness). Its a small Java program we wrote to use as a testing harness during opt compiler development. It allows you to force compilation of individual methods/classes in a fixed order and with different combintions of options.

--dave

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

[rvm-research] Detailed warnings

2009-11-20T02:15:32Z

Hi all,

I've been doing modifications to the JikesRVM and after the build it
reports the number of warnings, but I cannot find the details of the
warnings anywhere.
Where/how do I view the warnings themselves?

Thanks,

Luke Terry

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: Help in implementing on-the-fly gc

2009-11-19T23:34:43Z

The buffers i'm talking about are simple lists of ObjectReference s. There are many of them, Updates, youngObjects, roots, etc. By handshake, I trying to pause one mutator thread at a time and collect it's buffer values to the collector thread's buffers or change it's flags etc.

Thankyou for your reply, I now understand how the phases work. I am also kind of stuck on simple write barriers. I want to implement the copy-on-write mechanism. I tried to use the one in concMS but then figured it's already in this version. how should I use that one?

Re: [rvm-research] Opt compiler options not working

2009-11-19T21:15:28Z

If you are trying to write small test cases to see what the opt compiler is doing, you might want to learn about OptTestHarness (see the userguide, online at http://jikesrvm.org/OptTestHarness). Its a small Java program we wrote to use as a testing harness during opt compiler development. It allows you to force compilation of individual methods/classes in a fixed order and with different combintions of options.

--dave

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] Opt compiler options not working

2009-11-19T21:07:46Z

Try -X:opt:verbose=true to see if the opt compiler is actually compiling anything. My guess would be that your program is short running enough that the adaptive system never actually even invoked the opt compiler.

--dave

"srinivas(shankar)" <srinivas13.2@...> wrote on 11/19/2009 09:04:38 PM: > From:
> > "srinivas(shankar)" <srinivas13.2@...>
> > To:
> > "General discussion of Jikes RVM design, implementation, issues, and > plans" <jikesrvm-researchers@...>
> > Date:
> > 11/19/2009 09:05 PM
> > Subject:
> > Re: [rvm-research] Opt compiler options not working
> > The command i used was: > $rvm -X:opt:print_final_hir=true MyClass > here MyClass is a simple multi-threaded program intended to be a test case. > > The other one was: > $rvm -X:opt:escape_simple_ipa=true MyClass > > It just displays the ouput and quits - like the output obtained by > using plain $rvm MyClass without options.
> On Fri, Nov 20, 2009 at 10:22 AM, Robin Garner <robin.garner@... > > wrote:
> Srinivas, > > Could you send the full command line you're using to the list ?
> > srinivas13.2 wrote: > > hi all, > > None of the opt compiler options (like print_final_hir, print_final_lir etc) > > are giving me any output. Im unable to print the ir using these options. > > Also i included a VM.sysWriteln("Entering SimpleEscape.java..."); at the > > beginning of simpleEscapeAnalysis() in SimpleEscape.java and even > > explicitly set escape_simple_ipa=true but the line is not being printed on > > the screen, so i assume "escape_simple_ipa" isn't working as well. Is there > > any other flag that i dont know of which is to be set in order to perform > > escape analysis and escape transformations? > > Kindly help. > > > > Regards, > > Srinivas Gopal >
> ------------------------------------------------------------------------------ > Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day > trial. Simplify your report design, integration and deployment - and focus on > what you do best, core application coding. Discover what's new with > Crystal Reports now. http://p.sf.net/sfu/bobj-july > _______________________________________________ > Jikesrvm-researchers mailing list > Jikesrvm-researchers@... > https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers
> ------------------------------------------------------------------------------ > Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day > trial. Simplify your report design, integration and deployment - andfocus on > what you do best, core application coding. Discover what's new with > Crystal Reports now. http://p.sf.net/sfu/bobj-july > _______________________________________________ > Jikesrvm-researchers mailing list > Jikesrvm-researchers@... > https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers
------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

[rvm-research] Using PIN in Jikes RVM 3.0.0

2009-11-19T21:05:14Z

Hi,

I am trying to use PIN in Jikes RVM 3.0.0 (32bit application) on a 64-
bit Redhat-Linux OS.
I am a newbie for PIN. Can someone tell which version of PIN should I
download and what
command should I use to instrument in jikesrvm ? I downloaded the 64-
bit and used the following
command:

rvm -wrap "$HOME/tools/pin-2.6/ia32/bin/pinbin -t $HOME/tools/pin-2.6/
source/tools/SimpleExamples/obj-intel64/icount.so -- " Filename

I get the following error:
E:Unable to load $HOME/tools/pin-2.6/source/tools/SimpleExamples/obj-
intel64/icount.so

Best Regards,
Raj

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] Opt compiler options not working

2009-11-19T21:04:38Z

The command i used was:
$rvm -X:opt:print_final_hir=true MyClass
here MyClass is a simple multi-threaded program intended to be a test case.

The other one was:
$rvm -X:opt:escape_simple_ipa=true MyClass

It just displays the ouput and quits - like the output obtained by using plain $rvm MyClass without options.

On Fri, Nov 20, 2009 at 10:22 AM, Robin Garner <robin.garner@...> wrote:

Srinivas,

Could you send the full command line you're using to the list ?

srinivas13.2 wrote:
> hi all,
> None of the opt compiler options (like print_final_hir, print_final_lir etc)
> are giving me any output. Im unable to print the ir using these options.
> Also i included a VM.sysWriteln("Entering SimpleEscape.java..."); at the
> beginning of simpleEscapeAnalysis() in SimpleEscape.java and even
> explicitly set escape_simple_ipa=true but the line is not being printed on
> the screen, so i assume "escape_simple_ipa" isn't working as well. Is there
> any other flag that i dont know of which is to be set in order to perform
> escape analysis and escape transformations?
> Kindly help.
>
> Regards,
> Srinivas Gopal

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] Opt compiler options not working

2009-11-19T20:52:05Z

Srinivas,

Could you send the full command line you're using to the list ?

srinivas13.2 wrote:

> hi all,
> None of the opt compiler options (like print_final_hir, print_final_lir etc)
> are giving me any output. Im unable to print the ir using these options.
> Also i included a VM.sysWriteln("Entering SimpleEscape.java..."); at the
> beginning of simpleEscapeAnalysis() in SimpleEscape.java and even
> explicitly set escape_simple_ipa=true but the line is not being printed on
> the screen, so i assume "escape_simple_ipa" isn't working as well. Is there
> any other flag that i dont know of which is to be set in order to perform
> escape analysis and escape transformations?
> Kindly help.
>
> Regards,
> Srinivas Gopal

[rvm-research] questions for using PIN software

2009-11-19T20:48:56Z

Hi,
We are trying to use PIN (an emulator) to get some statistics from
running JikeRVM upon it. Does anyone use PIN to profiling JikesRVM on
64-linux system? Since the JikesRVM we are using is 3.0.0, so we don't
have 64-bits version.
Thanks,

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Opt compiler options not working

2009-11-19T20:41:41Z

hi all,
None of the opt compiler options (like print_final_hir, print_final_lir etc) are giving me any output. Im unable to print the ir using these options. Also i included a VM.sysWriteln("Entering SimpleEscape.java..."); at the beginning of simpleEscapeAnalysis() in SimpleEscape.java and even explicitly set escape_simple_ipa=true but the line is not being printed on the screen, so i assume "escape_simple_ipa" isn't working as well. Is there any other flag that i dont know of which is to be set in order to perform escape analysis and escape transformations?
Kindly help.

Regards,
Srinivas Gopal

Re: [rvm-research] What are the typical reasons if assertImmovable() fails

2009-11-19T16:56:37Z

lltong wrote:
> Dear All,
>
> What happens if the following stack trace is dumped?
> Thanks. By the way, I am using Jikes RVM 2.45, since I don't need the
> discontiguous mechanism.

Ouch. I hope you have a better reason than that for using 2.4.5 - it's
pretty old and a lot has changed since then. And you don't get the MMTk
test harness.

> Lcom/ibm/JikesRVM/VM; sysFail(Ljava/lang/String;)V at line 1079
> Lcom/ibm/JikesRVM/VM;
> _assertionFailure(Ljava/lang/String;Ljava/lang/String;)V at line 577
> Lcom/ibm/JikesRVM/VM; _assert(ZLjava/lang/String;Ljava/lang/String;)V
> at line 558
> Lcom/ibm/JikesRVM/VM; _assert(Z)V at line 538
> Lorg/mmtk/vm/ScanThread; assertImmovable()V at line 553
> Lorg/mmtk/vm/ScanThread;
> scanThreadInternal(Lorg/vmmagic/unboxed/Address;I)V at line 223
> Lorg/mmtk/vm/ScanThread;
> startScan(Lorg/mmtk/plan/TraceLocal;ZLcom/ibm/JikesRVM/VM_Thread;Lorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;)V
> at line 200
> Lorg/mmtk/vm/ScanThread;
> scanThread(Lcom/ibm/JikesRVM/VM_Thread;Lorg/mmtk/plan/TraceLocal;ZLorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;)V
> at line 166
> Lorg/mmtk/vm/ScanThread;
> scanThread(Lcom/ibm/JikesRVM/VM_Thread;Lorg/mmtk/plan/TraceLocal;Z)V at
> line 131
> Lorg/mmtk/vm/Scanning; computeAllRoots(Lorg/mmtk/plan/TraceLocal;)V
> at line 236
>

This stack trace says that while MMTk was scanning the stacks, it came
across an object that it thought should have been immovable, but which
wasn't. Either this object should have been copied by precopy, but
wasn't, or it should have been allocated immortal and wasn't. Read the
code for the method that threw the assertion, and then back down the
call stack to find out what's going on.

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

[rvm-research] What are the typical reasons if assertImmovable() fails

2009-11-19T05:58:29Z

Dear All,

What happens if the following stack trace is dumped?
Thanks. By the way, I am using Jikes RVM 2.45, since I don't need the
discontiguous mechanism.

Lcom/ibm/JikesRVM/VM; sysFail(Ljava/lang/String;)V at line 1079
Lcom/ibm/JikesRVM/VM;
_assertionFailure(Ljava/lang/String;Ljava/lang/String;)V at line 577
Lcom/ibm/JikesRVM/VM; _assert(ZLjava/lang/String;Ljava/lang/String;)V
at line 558
Lcom/ibm/JikesRVM/VM; _assert(Z)V at line 538
Lorg/mmtk/vm/ScanThread; assertImmovable()V at line 553
Lorg/mmtk/vm/ScanThread;
scanThreadInternal(Lorg/vmmagic/unboxed/Address;I)V at line 223
Lorg/mmtk/vm/ScanThread;
startScan(Lorg/mmtk/plan/TraceLocal;ZLcom/ibm/JikesRVM/VM_Thread;Lorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;)V
at line 200
Lorg/mmtk/vm/ScanThread;
scanThread(Lcom/ibm/JikesRVM/VM_Thread;Lorg/mmtk/plan/TraceLocal;ZLorg/vmmagic/unboxed/Address;Lorg/vmmagic/unboxed/Address;)V
at line 166
Lorg/mmtk/vm/ScanThread;
scanThread(Lcom/ibm/JikesRVM/VM_Thread;Lorg/mmtk/plan/TraceLocal;Z)V at
line 131
Lorg/mmtk/vm/Scanning; computeAllRoots(Lorg/mmtk/plan/TraceLocal;)V
at line 236

--
Regards,
Liangliang

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] Help in implementing on-the-fly gc

2009-11-18T22:38:45Z

lalli_maven wrote:

> I've solved the first problem by modifying the abstract classes. But still
> stuck on the other three. Please help.
>
> lalli_maven wrote:
>> I'm trying to implement an on-the-fly gc, which is mostly on the lines of
>> the age oriented gc. I'm stuck on a lot of issues:
>>
>> 1. The phases of the gc have to be implemented in the MMTk/src group, but
>> the handshakes have to be implemented in the RVMThread.java which is in
>> rvm/src (not visible in MMTk/src), how will that work out?
>>
>> 2. The buffer variables are "declared" in the collector code in the MMTk
>> package, for which I also modified Selected.java file, but I kind of
>> messed up everything. Where am I supposed to declare these variables?

Which buffer variables ? What are you trying to achieve by modifying them ?

>> 3. The way any gc works is such a way that it pushes a collection phase
>> into a stack then pops it and executes it. Why? The docs of Phase.java (in
>> MMTk/src) indicate that any handshaking we have to do is to be done here
>> itself. How?

This was/is part of recasting the MMTk phase scheduler as a state
machine so that a concurrent GC phase can perform an increment of GC
then save and resume where it left off. The scheduling algorithm for
ConcMS was:

1. Identify that a GC should be triggered.

2. Stop the world, enumerate roots and start the concurrent mark thread.

3. When either a) the concurrent mark thread has finished, or b) an
out-of-memory condition occurs

4. Stop the world, complete the mark phase (in parallel) and finish the GC.

So from state (3), there is a transition to either state (3) or (4),
requiring some mechanism for different threads to pick up the GC state
and continue processing. Hence the phase stack.

Feel free to replace the phase stack with something more elegant :)

The term 'handshake' is used in different ways in different places in
the code. What do you mean by 'handshake' ?

>> 4. The run function in CollectorThread.java is very complicated. It's
>> supposed to only do a hardhandshake-suspend all other threads, collect and
>> then hardhandshake-resume all other threads. But what it is actually doing
>> is calling many rendezvous functions with pointless integer parameters.
>> What is that about? What I want to do is simply keep the parking function
>> and keep the loop inside which only collect is called. How can I do that?

The integer parameters are not pointless. They are used by error
reporting code, so that when JikesRVM hangs at a rendezvous point, the
error message tells you where the problem was. They are also used to
ensure that two threads don't attempt to rendezvous at different points.

The ConcMS collector had a concurrent mark phase with two brief
stop-the-world phases at the beginning and end of every GC cycle. Doing
true on-the-fly GC in JikesRVM will involve reimplementing much of the
code you are referring to. It's not a small job, and a difficult one
even for people who have been working with JikesRVM for many years.

Regards,
Robin

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: Help in implementing on-the-fly gc

2009-11-18T19:54:28Z

I've solved the first problem by modifying the abstract classes. But still stuck on the other three. Please help.

lalli_maven wrote:

I'm trying to implement an on-the-fly gc, which is mostly on the lines of the age oriented gc. I'm stuck on a lot of issues:

1. The phases of the gc have to be implemented in the MMTk/src group, but the handshakes have to be implemented in the RVMThread.java which is in rvm/src (not visible in MMTk/src), how will that work out?

2. The buffer variables are "declared" in the collector code in the MMTk package, for which I also modified Selected.java file, but I kind of messed up everything. Where am I supposed to declare these variables?

3. The way any gc works is such a way that it pushes a collection phase into a stack then pops it and executes it. Why? The docs of Phase.java (in MMTk/src) indicate that any handshaking we have to do is to be done here itself. How?

4. The run function in CollectorThread.java is very complicated. It's supposed to only do a hardhandshake-suspend all other threads, collect and then hardhandshake-resume all other threads. But what it is actually doing is calling many rendezvous functions with pointless integer parameters. What is that about? What I want to do is simply keep the parking function and keep the loop inside which only collect is called. How can I do that?

Please help me!

Re: [rvm-research] Problems in Partitioning for implementing Connectivity Based Garbage Collector!!

2009-11-18T16:37:21Z

maverick.vickey wrote:
> Hi Robin!!!
> Actually I need to implement partitions that include classes & objects in
> bootimage, I don need to find out if my classes are there in bootimage or
> not!! I need a way to retrieve classes & objects in bootimage at some other
> place, say package "org.jikesrvm"!!
> Rahul Bhardwaj
> Postgraduation Student
> Computer Science
> BITS-Pilani, India

Currently the boot image is mapped as a single immortal space, and is
treated by GC as a sparse array of roots. If you want to change the way
the boot image is laid out, you will need to look at the class
BootImageWriter (in tools/bootImageWriter/src).

If you want to scan the boot image during GC rather than treating it as
roots, you should be able to change the constant Plan.SCAN_BOOT_IMAGE to
false, but you may well need to do some debugging as this is an untested
configuration, and has been for a few years.

If you want to change the layout of the boot image but still treat it as
a source of roots you'll need to consider the class
org.jikesrvm.mm.mmtk.ScanBootImage.

hth,
Robin

>
> Robin Garner wrote:
>> maverick.vickey wrote:
>>> I am trying to implement Connectivity Based Garbage Collector. I am able
>>> to
>>> implement Partitioning for dynamically loaded classes but I am unable to
>>> include classes & objects in bootimage into the partitions!!
>>> Please help!!!
>> Please describe the problem in more detail. What classes have you
>> created ? How do you know they don't show up in the boot image ?
>>
>> And could you please include your affiliation when posting to this list.
>> Thanks,
>> Robin
>>
>> --
>> Robin Garner
>> PhD Student
>> Australian National University
>>
>> ------------------------------------------------------------------------------
>> Let Crystal Reports handle the reporting - Free Crystal Reports 2008
>> 30-Day
>> trial. Simplify your report design, integration and deployment - and focus
>> on
>> what you do best, core application coding. Discover what's new with
>> Crystal Reports now. http://p.sf.net/sfu/bobj-july
>> _______________________________________________
>> Jikesrvm-researchers mailing list
>> Jikesrvm-researchers@...
>> https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers
>>
>>
>

not able to import jikes files in test cases properly

2009-11-18T09:44:45Z

hi,
i m tryin to call some jikes files for example Memorymanager.java imto my test cases using proper import statements and rules and since i am using eclipse when i place a dot with file name it shows all options(methods)usable but after tryin to run i get error like nullpointer exception and i have no idea to get over this problem??

Re: [rvm-research] Problems in Partitioning for implementing Connectivity Based Garbage Collector!!

2009-11-17T01:48:20Z

Hi Robin!!!
Actually I need to implement partitions that include classes & objects in bootimage, I don need to find out if my classes are there in bootimage or not!! I need a way to retrieve classes & objects in bootimage at some other place, say package "org.jikesrvm"!!
Rahul Bhardwaj
Postgraduation Student
Computer Science
BITS-Pilani, India

Robin Garner wrote:

maverick.vickey wrote:
> I am trying to implement Connectivity Based Garbage Collector. I am able to
> implement Partitioning for dynamically loaded classes but I am unable to
> include classes & objects in bootimage into the partitions!!
> Please help!!!

Please describe the problem in more detail. What classes have you
created ? How do you know they don't show up in the boot image ?

And could you please include your affiliation when posting to this list.
Thanks,
Robin

--
Robin Garner
PhD Student
Australian National University

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] Problems in Partitioning for implementing Connectivity Based Garbage Collector!!

2009-11-16T22:26:12Z

maverick.vickey wrote:
> I am trying to implement Connectivity Based Garbage Collector. I am able to
> implement Partitioning for dynamically loaded classes but I am unable to
> include classes & objects in bootimage into the partitions!!
> Please help!!!

Please describe the problem in more detail. What classes have you
created ? How do you know they don't show up in the boot image ?

And could you please include your affiliation when posting to this list.
Thanks,
Robin

--
Robin Garner
PhD Student
Australian National University

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] Unable to modify the code base (OutOfMemory error)

2009-11-16T13:03:23Z

Ok here it is, when i add a VM.sysWriteln in VM.java i do not get any error, but if i try to change the value of DEBUG to true in the method simpleEscapeAnalysis() in SimpleEscape.java or for that matter, add the same VM.sysWrite() there, the build process fails giving the following message:

BUILD FAILED
/home/shankar/jikesrvm-3.1.0/build.xml:1434: The following error occurred while executing this line:
/home/shankar/jikesrvm-3.1.0/build.xml:1441: java.lang.OutOfMemoryError: Java heap space

Im not using svn.

Also the even after using the "print_final_hir" option, rvm does not print the HIR to the screen.Is there another way to do the same?

Thanks,
Srinivas Gopal

Robin Garner wrote:

srinivas13.2 wrote:
> hi all,
> Whenever i try to make any modifications in the code base, i get OutOfMemory
> error while building the rvm (even if the modification is something as small
> as adding a VM.sysWrite()). The problem is similar to the one described in
> http://old.nabble.com/outofmemory--error-on-building-to25643001.html
> outofmemory error on building
> So, as per the given instructions i have increased the value of maxmemory in
> build.xml from 500M (default) to 2000M. Even then it doesn't seem to help.
> Kindly let me know what is to be done.
>
> Regards,
> Srinivas Gopal

I expect the problem is specific to the change you are making.

First, try to reduce the problem to the smallest change, eg the simple
VM.syswrite you refer to above, then do an svn diff, and post the output
here.

cheers,
Robin

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Problems in Partitioning for implementing Connectivity Based Garbage Collector!!

2009-11-16T10:44:12Z

I am trying to implement Connectivity Based Garbage Collector. I am able to implement Partitioning for dynamically loaded classes but I am unable to include classes & objects in bootimage into the partitions!!
Please help!!!

Re: [rvm-research] how to make a class visible in different package

2009-11-16T09:34:06Z

Suriya Subramanian wrote:
> On Mon, Nov 16, 2009 at 01:21:47AM -0800, dsharma wrote:
>> How can i make class files in rvm/src visible in MMTk/src.
>
> You cannot do so in full generality. The interface between the RVM and MMTk
> is very narrow with the VM exposing only what MMTk needs. Take a look at
> the MMTk user guide (Appendix A, VM Interface Specification) and at the
> source code MMTk/ext/vm/jikesrvm to see how is interface looks.

Fil Pizlo had a nice post a little while back on the specifics of the
RVM-MMTk interface:
http://old.nabble.com/Re%3A--rvm-research--Making-Stop-the-world-on-the-fly-p25898295.html

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] how to make a class visible in different package

2009-11-16T09:11:45Z

On Mon, Nov 16, 2009 at 01:21:47AM -0800, dsharma wrote:
>
> How can i make class files in rvm/src visible in MMTk/src.

You cannot do so in full generality. The interface between the RVM and MMTk
is very narrow with the VM exposing only what MMTk needs. Take a look at
the MMTk user guide (Appendix A, VM Interface Specification) and at the
source code MMTk/ext/vm/jikesrvm to see how is interface looks.

Suriya

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

[rvm-research] how to make a class visible in different package

2009-11-16T01:21:47Z

How can i make class files in rvm/src visible in MMTk/src.
--
View this message in context: http://old.nabble.com/how-to-make-a-class-visible-in-different-package-tp26368936p26368936.html
Sent from the jikesrvm-researchers mailing list archive at Nabble.com.

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Re: [rvm-research] Unable to modify the code base (OutOfMemory error)

2009-11-15T15:49:05Z

srinivas13.2 wrote:

> hi all,
> Whenever i try to make any modifications in the code base, i get OutOfMemory
> error while building the rvm (even if the modification is something as small
> as adding a VM.sysWrite()). The problem is similar to the one described in
> http://old.nabble.com/outofmemory--error-on-building-to25643001.html
> outofmemory error on building
> So, as per the given instructions i have increased the value of maxmemory in
> build.xml from 500M (default) to 2000M. Even then it doesn't seem to help.
> Kindly let me know what is to be done.
>
> Regards,
> Srinivas Gopal

I expect the problem is specific to the change you are making.

First, try to reduce the problem to the smallest change, eg the simple
VM.syswrite you refer to above, then do an svn diff, and post the output
here.

cheers,
Robin

------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Jikesrvm-researchers mailing list
Jikesrvm-researchers@...
https://lists.sourceforge.net/lists/listinfo/jikesrvm-researchers

Help in implementing on-the-fly gc

2009-11-15T10:59:08Z

I'm trying to implement an on-the-fly gc, which is mostly on the lines of the age oriented gc. I'm stuck on a lot of issues:

1. The phases of the gc have to be implemented in the MMTk/src group, but the handshakes have to be implemented in the RVMThread.java which is in rvm/src (not visible in MMTk/src), how will that work out?

2. The buffer variables are "declared" in the collector code in the MMTk package, for which I also modified Selected.java file, but I kind of messed up everything. Where am I supposed to declare these variables?

3. The way any gc works is such a way that it pushes a collection phase into a stack then pops it and executes it. Why? The docs of Phase.java (in MMTk/src) indicate that any handshaking we have to do is to be done here itself. How?

4. The run function in CollectorThread.java is very complicated. It's supposed to only do a hardhandshake-suspend all other threads, collect and then hardhandshake-resume all other threads. But what it is actually doing is calling many rendezvous functions with pointless integer parameters. What is that about? What I want to do is simply keep the parking function and keep the loop inside which only collect is called. How can I do that?

Please help me!