[mpich-discuss] MCS lock and MPI RMA problem
Ask Jakobsen
afj at qeye-labs.com
Mon Mar 13 13:43:24 CDT 2017
Thanks Pavan and Halim. You are right it progress in the Fetch_and_op
version without the async progress environment variable. I will try to
implement the MPI_MODE_NOCHECK as you suggested.
To make matters more complicated:
I have discovered that the code from the book in mcs-lock.c deviates
from "High-Performance
Distributed RMA Locks" pseudo code (see Listing 3 in paper) and the
original MCS paper "Algorithms for scalable Sync on shared memory
multiprocessors". If I add to the original mcs-lock.c code
lmem[nextRank]=-1;
before entering the MPI_win_lock_all in acquire the code *almost appears*
to work! Sort of... when having a large number of processes there are still
a rare race condition where a few processes don't get to the MPI_Win_free(&win)
in main().
On Mon, Mar 13, 2017 at 6:15 PM, Halim Amer <aamer at anl.gov> wrote:
> To be precise, asynchronous progress is not required for this second
> implementation because the busy waiting loop is doing a Fetch_and_op. It is
> required, however, for the first implementation, from the tutorial book,
> because it busy waits with Win_sync.
>
> Halim
> www.mcs.anl.gov/~aamer
>
>
> On 3/13/17 8:46 AM, Balaji, Pavan wrote:
>
>>
>> I should also point out that I don't think your implementation is
>> assuming asynchronous progress. You shouldn't have to do any of the
>> asynchronous progress tweaks for it to work correctly.
>>
>> -- Pavan
>>
>> On Mar 13, 2017, at 8:43 AM, Balaji, Pavan <balaji at anl.gov> wrote:
>>>
>>>
>>> OK, I spent a little more time going through the code. The algorithm
>>> looks correct, except for some minor issues:
>>>
>>> 1. mcs-lock-fop.c:72 -- you need a flush or flush_local. You were lucky
>>> that this was working correctly since it's local, but the MPI standard
>>> doesn't guarantee it.
>>>
>>> 2. You might be able to simplify mcs-lock-fop.c lines 72-90 as follows:
>>>
>>> do {
>>> MPI_Fetch_and_op(&dummy, &fetch_nextrank, MPI_INT,
>>> myrank, nextRank, MPI_NO_OP, win);
>>> MPI_Win_flush(myrank, win);
>>> } while (fetch_nextrank==-1);
>>>
>>> 3. Polling on the nextrank value is better than polling on a remote
>>> location. However, you could further simplify this by using send/recv to
>>> notify the waiting process rather than RMA. This allows the MPI
>>> implementation the opportunity to block waiting for progress, rather than
>>> poll (though in practice, current implementations poll anyway).
>>>
>>> 4. Since you are always using the lock in shared mode, you should
>>> specify the hint MPI_MODE_NOCHECK in your lock_all epochs.
>>>
>>> Now, coming to your bug, this does seem to be a bug in the MPI
>>> implementation. We can dig into it further. In the meanwhile, if you use
>>> the optimization #4 above, this will allow the MPI implementation to bypass
>>> the entire locking checks, which will get you past the bug for now.
>>>
>>> Thanks for reporting the issue.
>>>
>>> -- Pavan
>>>
>>> On Mar 13, 2017, at 3:06 AM, Ask Jakobsen <afj at qeye-labs.com> wrote:
>>>>
>>>> I don't think so. Rank 0 also holds the tail which is the process which
>>>> most recently requested the mutex.
>>>>
>>>> On Mon, Mar 13, 2017 at 2:55 AM, Balaji, Pavan <balaji at anl.gov> wrote:
>>>>
>>>> Shouldn't winsize be 3 integers in your code? (sorry, I spent only 30
>>>> seconds looking at the code, so I might have missed something).
>>>>
>>>> -- Pavan
>>>>
>>>> On Mar 12, 2017, at 2:44 PM, Ask Jakobsen <afj at qeye-labs.com> wrote:
>>>>>
>>>>> Interestingly, according to the paper you suggested it appears to
>>>>> include a similar test in pseudo code https://htor.inf.ethz.ch/publi
>>>>> cations/img/hpclocks.pdf (see Listing 3 in paper).
>>>>>
>>>>> Unfortunately, removing the test in the release protocol did not solve
>>>>> the problem. The race condition is much more difficult to provoke, but I
>>>>> managed when setting the size of the communicator to 3 (only tested even
>>>>> sizes so far).
>>>>>
>>>>> From Jeff's suggestion I have attempted to rewrite the code removing
>>>>> local loads and stores in the MPI_Win_lock_all epochs using
>>>>> MPI_Fetch_and_op (see attached files).
>>>>>
>>>>> This version behaves very similar to the original code and also fails
>>>>> from time to time. Putting a sleep into the acquire busy loop (usleep(100))
>>>>> will make the code "much more robust" (I hack, I know, but indicating some
>>>>> underlying race condition?!). Let me know if you see any problems in the
>>>>> way I am using MPI_Fetch_and_op in a busy loop. Flushing or syncing is not
>>>>> necessary in this case, right?
>>>>>
>>>>> All work is done with export MPIR_CVAR_ASYNC_PROGRESS=1 on mpich-3.2
>>>>> and mpich-3.3a2
>>>>>
>>>>> On Wed, Mar 8, 2017 at 4:21 PM, Halim Amer <aamer at anl.gov> wrote:
>>>>> I cannot claim that I thoroughly verified the correctness of that
>>>>> code, so take it with a grain of salt. Please keep in mind that it is a
>>>>> test code from a tutorial book; those codes are meant for learning purposes
>>>>> not for deployment.
>>>>>
>>>>> If your goal is to have a high performance RMA lock, I suggest you to
>>>>> look into the recent HPDC'16 paper: "High-Performance Distributed RMA
>>>>> Locks".
>>>>>
>>>>> Halim
>>>>> www.mcs.anl.gov/~aamer
>>>>>
>>>>> On 3/8/17 3:06 AM, Ask Jakobsen wrote:
>>>>> You are absolutely correct, Halim. Removing the test lmem[nextRank] ==
>>>>> -1
>>>>> in release fixes the problem. Great work. Now I will try to understand
>>>>> why
>>>>> you are right. I hope the authors of the book will credit you for
>>>>> discovering the bug.
>>>>>
>>>>> So in conclusion you need to remove the above mentioned test AND enable
>>>>> asynchronous progression using the environment variable
>>>>> MPIR_CVAR_ASYNC_PROGRESS=1 in MPICH (BTW I still can't get the code to
>>>>> work
>>>>> in openmpi).
>>>>>
>>>>> On Tue, Mar 7, 2017 at 5:37 PM, Halim Amer <aamer at anl.gov> wrote:
>>>>>
>>>>> detect that another process is being or already enqueued in the MCS
>>>>> queue.
>>>>>
>>>>> Actually the problem occurs only when the waiting process already
>>>>> enqueued
>>>>> itself, i.e., the accumulate operation on the nextRank field succeeded.
>>>>>
>>>>> Halim
>>>>> www.mcs.anl.gov/~aamer <http://www.mcs.anl.gov/%7Eaamer>
>>>>>
>>>>>
>>>>> On 3/7/17 10:29 AM, Halim Amer wrote:
>>>>>
>>>>> In the Release protocol, try removing this test:
>>>>>
>>>>> if (lmem[nextRank] == -1) {
>>>>> If-Block;
>>>>> }
>>>>>
>>>>> but keep the If-Block.
>>>>>
>>>>> The hang occurs because the process releasing the MCS lock fails to
>>>>> detect that another process is being or already enqueued in the MCS
>>>>> queue.
>>>>>
>>>>> Halim
>>>>> www.mcs.anl.gov/~aamer <http://www.mcs.anl.gov/%7Eaamer>
>>>>>
>>>>>
>>>>> On 3/7/17 6:43 AM, Ask Jakobsen wrote:
>>>>>
>>>>> Thanks, Halim. I have now enabled asynchronous progress in MPICH (can't
>>>>> find something similar in openmpi) and now all ranks acquire the lock
>>>>> and
>>>>> the program finish as expected. However if I put a while(1) loop
>>>>> around the
>>>>> acquire-release code in main.c it will fail again at random and go
>>>>> into an
>>>>> infinite loop. The simple unfair lock does not have this problem.
>>>>>
>>>>> On Tue, Mar 7, 2017 at 12:44 AM, Halim Amer <aamer at anl.gov> wrote:
>>>>>
>>>>> My understanding is that this code assumes asynchronous progress.
>>>>> An example of when the processes hang is as follows:
>>>>>
>>>>> 1) P0 Finishes MCSLockAcquire()
>>>>> 2) P1 is busy waiting in MCSLockAcquire() at
>>>>> do {
>>>>> MPI_Win_sync(win);
>>>>> } while (lmem[blocked] == 1);
>>>>> 3) P0 executes MCSLockRelease()
>>>>> 4) P0 waits on MPI_Win_lock_all() inside MCSLockRlease()
>>>>>
>>>>> Hang!
>>>>>
>>>>> For P1 to get out of the loop, P0 has to get out of
>>>>> MPI_Win_lock_all() and
>>>>> executes its Compare_and_swap().
>>>>>
>>>>> For P0 to get out MPI_Win_lock_all(), it needs an ACK from P1 that it
>>>>> got
>>>>> the lock.
>>>>>
>>>>> P1 does not make communication progress because MPI_Win_sync is not
>>>>> required to do so. It only synchronizes private and public copies.
>>>>>
>>>>> For this hang to disappear, one can either trigger progress manually by
>>>>> using heavy-duty synchronization calls instead of Win_sync (e.g.,
>>>>> Win_unlock_all + Win_lock_all), or enable asynchronous progress.
>>>>>
>>>>> To enable asynchronous progress in MPICH, set the
>>>>> MPIR_CVAR_ASYNC_PROGRESS
>>>>> env var to 1.
>>>>>
>>>>> Halim
>>>>> www.mcs.anl.gov/~aamer <http://www.mcs.anl.gov/%7Eaamer> <
>>>>> http://www.mcs.anl.gov/%7Eaamer>
>>>>>
>>>>>
>>>>> On 3/6/17 1:11 PM, Ask Jakobsen wrote:
>>>>>
>>>>> I am testing on x86_64 platform.
>>>>>
>>>>> I have tried to built both the mpich and the mcs lock code with -O0 to
>>>>> avoid agressive optimization. After your suggestion I have also
>>>>> tried to
>>>>> make volatile int *pblocked pointing to lmem[blocked] in the
>>>>> MCSLockAcquire
>>>>> function and volatile int *pnextrank pointing to lmem[nextRank] in
>>>>> MCSLockRelease, but it does not appear to make a difference.
>>>>>
>>>>> On suggestion from Richard Warren I have also tried building the code
>>>>> using
>>>>> openmpi-2.0.2 without any luck (however it appears to acquire the
>>>>> lock a
>>>>> couple of extra times before failing) which I find troubling.
>>>>>
>>>>> I think I will give up using local load/stores and will see if I can
>>>>> figure
>>>>> out if rewrite using MPI calls like MPI_Fetch_and_op as you suggest.
>>>>> Thanks for your help.
>>>>>
>>>>> On Mon, Mar 6, 2017 at 7:20 PM, Jeff Hammond <jeff.science at gmail.com>
>>>>> wrote:
>>>>>
>>>>> What processor architecture are you testing?
>>>>>
>>>>>
>>>>> Maybe set lmem to volatile or read it with MPI_Fetch_and_op rather
>>>>> than a
>>>>> load. MPI_Win_sync cannot prevent the compiler from caching *lmem
>>>>> in a
>>>>> register.
>>>>>
>>>>> Jeff
>>>>>
>>>>> On Sat, Mar 4, 2017 at 12:30 AM, Ask Jakobsen <afj at qeye-labs.com>
>>>>> wrote:
>>>>>
>>>>> Hi,
>>>>>
>>>>>
>>>>> I have downloaded the source code for the MCS lock from the excellent
>>>>> book "Using Advanced MPI" from http://www.mcs.anl.gov/researc
>>>>> h/projects/mpi/usingmpi/examples-advmpi/rma2/mcs-lock.c
>>>>>
>>>>> I have made a very simple piece of test code for testing the MCS lock
>>>>> but
>>>>> it works at random and often never escapes the busy loops in the
>>>>> acquire
>>>>> and release functions (see attached source code). The code appears
>>>>> semantically correct to my eyes.
>>>>>
>>>>> #include <stdio.h>
>>>>> #include <mpi.h>
>>>>> #include "mcs-lock.h"
>>>>>
>>>>> int main(int argc, char *argv[])
>>>>> {
>>>>> MPI_Win win;
>>>>> MPI_Init( &argc, &argv );
>>>>>
>>>>> MCSLockInit(MPI_COMM_WORLD, &win);
>>>>>
>>>>> int rank, size;
>>>>> MPI_Comm_rank(MPI_COMM_WORLD, &rank);
>>>>> MPI_Comm_size(MPI_COMM_WORLD, &size);
>>>>>
>>>>> printf("rank: %d, size: %d\n", rank, size);
>>>>>
>>>>>
>>>>> MCSLockAcquire(win);
>>>>> printf("rank %d aquired lock\n", rank); fflush(stdout);
>>>>> MCSLockRelease(win);
>>>>>
>>>>>
>>>>> MPI_Win_free(&win);
>>>>> MPI_Finalize();
>>>>> return 0;
>>>>> }
>>>>>
>>>>>
>>>>> I have tested on several hardware platforms and mpich-3.2 and
>>>>> mpich-3.3a2
>>>>> but with no luck.
>>>>>
>>>>> It appears that the MPI_Win_Sync are not "refreshing" the local
>>>>> data or
>>>>> I
>>>>> have a bug I can't spot.
>>>>>
>>>>> A simple unfair lock like http://www.mcs.anl.gov/researc
>>>>> h/projects/mpi/usingmpi/examples-advmpi/rma2/ga_mutex1.c works
>>>>> perfectly.
>>>>>
>>>>> Best regards, Ask Jakobsen
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> Jeff Hammond
>>>>> jeff.science at gmail.com
>>>>> http://jeffhammond.github.io/
>>>>>
>>>>> _______________________________________________
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>>
>>>>> _______________________________________________
>>>>>
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>>
>>>>> _______________________________________________
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>>
>>>>> _______________________________________________
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>>
>>>>> _______________________________________________
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>> <main.c><mcs-lock-fop.c><mcs-lock.h>________________________
>>>>> _______________________
>>>>> discuss mailing list discuss at mpich.org
>>>>> To manage subscription options or unsubscribe:
>>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>>
>>>>
>>>> _______________________________________________
>>>> discuss mailing list discuss at mpich.org
>>>> To manage subscription options or unsubscribe:
>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>
>>>>
>>>>
>>>> --
>>>> Ask Jakobsen
>>>> R&D
>>>>
>>>> Qeye Labs
>>>> Lersø Parkallé 107
>>>> 2100 Copenhagen Ø
>>>> Denmark
>>>>
>>>> mobile: +45 2834 6936
>>>> email: afj at Qeye-Labs.com
>>>> _______________________________________________
>>>> discuss mailing list discuss at mpich.org
>>>> To manage subscription options or unsubscribe:
>>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>>
>>>
>>> _______________________________________________
>>> discuss mailing list discuss at mpich.org
>>> To manage subscription options or unsubscribe:
>>> https://lists.mpich.org/mailman/listinfo/discuss
>>>
>>
>> _______________________________________________
>> discuss mailing list discuss at mpich.org
>> To manage subscription options or unsubscribe:
>> https://lists.mpich.org/mailman/listinfo/discuss
>>
>> _______________________________________________
> discuss mailing list discuss at mpich.org
> To manage subscription options or unsubscribe:
> https://lists.mpich.org/mailman/listinfo/discuss
>
--
*Ask Jakobsen*
R&D
Qeye Labs
Lersø Parkallé 107
2100 Copenhagen Ø
Denmark
mobile: +45 2834 6936
email: afj at Qeye-Labs.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20170313/6cbbd68f/attachment.html>
More information about the discuss
mailing list