On Tue, 2017-02-28 at 17:59 +0000, Giuseppe Congiu wrote:

Hello,

I am trying to integrate an asynchronous API in ROMIO but I am experiencing some problems, probably caused by the improper use of the MPIX_Grequest interface.

I'm excited to see someone using the MPIX_Grequest interface. We used the MPIX_Grequest interface to implement non-blocking collective I/O, and had some bad interactions between libc's aio and the grequest callbacks. I don't know if you are running into something similar.

Short description of the integrated functions: I have a ROMIO implementation with extended MPI-IO hints that can be used to divert written data, during collective I/O, to a local file system (typically running on a SSD). The full description of the modification can be found at this link: https://github.com/gcongiu/E10/tree/beegfs-devel

Do you have any desire or plans to submit these changes into upstream ROMIO? There have been quite a few changes since 3.1.4, particularly for dealing with large transfers. I don't think you're hitting an overflow path here, but it's something I cannot rule out.

The BeeGFS ROMIO driver uses the cache API internally provided by BeeGFS to move data around, while the custom ROMIO implementation provides support for any other file system (in the "common" layer).

Description of the problem: I am currently having problems making the BeeGFS driver work properly. More specifically I am noticing that when the size of the shared file is large (32GB in my configuration) and I am writing multiple files one after the other (using a custom coll_perf benchmark) the i-th file (for i > 0 with 0 <= i < 4) gets stuck.

BeeGFS provides a deeper_cache_flush_range() function to flush data in the cache to the global file. This is non blocking and will submit a range (offset, length) for a certain filename to a cache deamon that transfers the data. Multiple ranges can be submitted one after the other. 

Completion can be checked using deeper_cache_flush_wait() using the filename as input. Since the MPI_Grequest model requires the external thread to make progress by invoking MPI_Grequest_complete() I have opted for the MPIX_Grequest interface which allows me to make progress (MPI_Grequest_complete) inside MPI while invoking the MPI_Wait() function.

The original intent of the MPIX_Grequest extensions is that something else (in this case the AIO library) was making progress in the background and all we needed was a way to check if things were done or not. I suspect the callbacks could do some work, but blocking callbacks haven't seen a lot of testing (in 10+ years the extensions themselves haven't seen a lot of activity, either)

deeper_cache_flush_range() does not require the application to keep any handler for the submitted request. The only thing the application needs is the name of the file to pass to deeper_cache_flush_wait().

Follows a code snippet for the corresponding non blocking implementation in ROMIO:

/* arguments for callback function */ struct callback {     ADIO_File fd_;     ADIOI_Sync_req_t req_; };

/*  * ADIOI_BEEGFS_Sync_thread_start - start synchronisation of req  */ int ADIOI_BEEGFS_Sync_thread_start(ADIOI_Sync_thread_t t) {     ADIOI_Atomic_queue_t q = t->sub_;     ADIOI_Sync_req_t r;     int retval, count, fflags, error_code, i;     ADIO_Offset offset, len;     MPI_Count datatype_size;     MPI_Datatype datatype;     ADIO_Request *req;     char myname[] = "ADIOI_BEEGFS_SYNC_THREAD_START";

    r = ADIOI_Atomic_queue_front(q);     ADIOI_Atomic_queue_pop(q);

    ADIOI_Sync_req_get_key(r, ADIOI_SYNC_ALL, &offset,    &datatype, &count, &req, &error_code, &fflags);

    MPI_Type_size_x(datatype, &datatype_size);     len = (ADIO_Offset)datatype_size * (ADIO_Offset)count;

    retval = deeper_cache_flush_range(t->fd_->filename, (off_t)offset, (size_t)len, fflags);

    if (retval == DEEPER_RETVAL_SUCCESS && ADIOI_BEEGFS_greq_class == 1) { MPIX_Grequest_class_create(ADIOI_BEEGFS_Sync_req_query,                    ADIOI_BEEGFS_Sync_req_free,                    MPIU_Greq_cancel_fn,            ADIOI_BEEGFS_Sync_req_poll,            ADIOI_BEEGFS_Sync_req_wait,    &ADIOI_BEEGFS_greq_class);     } else { /* --BEGIN ERROR HANDLING-- */ return MPIO_Err_create_code(MPI_SUCCESS,                     MPIR_ERR_RECOVERABLE,                     "ADIOI_BEEGFS_Cache_sync_req",             __LINE__, MPI_ERR_IO, "**io %s",             strerror(errno));         /* --END ERROR HANDLING-- */     }

    /* init args for the callback functions */     struct callback *args = (struct callback *)ADIOI_Malloc(sizeof(struct callback));     args->fd_ = t->fd_;     args->req_ = r;

    MPIX_Grequest_class_allocate(ADIOI_BEEGFS_greq_class, args, req);

    return MPI_SUCCESS; }

/*  * ADIOI_BEEGFS_Sync_req_poll -  */ int ADIOI_BEEGFS_Sync_req_poll(void *extra_state, MPI_Status *status) {     struct callback *cb = (struct callback *)extra_state;     ADIOI_Sync_req_t r = (ADIOI_Sync_req_t)cb->req_;     ADIO_File fd = (ADIO_File)cb->fd_;     char *filename = fd->filename;     int count, cache_flush_flags, error_code;     MPI_Datatype datatype;     ADIO_Offset offset;     MPI_Aint lb, extent;     ADIO_Offset len;     ADIO_Request *req;

    ADIOI_Sync_req_get_key(r, ADIOI_SYNC_ALL, &offset,    &datatype, &count, &req, &error_code, &cache_flush_flags);

    int retval = deeper_cache_flush_wait(filename, cache_flush_flags);

    MPI_Type_get_extent(datatype, &lb, &extent);     len = (ADIO_Offset)extent * (ADIO_Offset)count;

    if (fd->hints->e10_cache_coherent == ADIOI_HINT_ENABLE) ADIOI_UNLOCK(fd, offset, SEEK_SET, len);

    /* mark generilized request as completed */     MPI_Grequest_complete(*req);

    if (retval != DEEPER_RETVAL_SUCCESS)         goto fn_exit_error;

    MPI_Status_set_cancelled(status, 0);     MPI_Status_set_elements(status, datatype, count);     status->MPI_SOURCE = MPI_UNDEFINED;     status->MPI_TAG = MPI_UNDEFINED;

    ADIOI_Free(cb);

    return MPI_SUCCESS;

fn_exit_error:     ADIOI_Free(cb);

    return MPIO_Err_create_code(MPI_SUCCESS,                                 MPIR_ERR_RECOVERABLE,                                 "ADIOI_BEEGFS_Sync_req_poll",                                 __LINE__, MPI_ERR_IO, "**io %s",                                 strerror(errno)); }

/*  * ADIOI_BEEGFS_Sync_req_wait -  */ int ADIOI_BEEGFS_Sync_req_wait(int count, void **array_of_states, double timeout, MPI_Status *status) {     return ADIOI_BEEGFS_Sync_req_poll(*array_of_states, status); }

/*  * ADIOI_BEEGFS_Sync_req_query -  */ int ADIOI_BEEGFS_Sync_req_query(void *extra_state, MPI_Status *status) {     return MPI_SUCCESS; }

/*  * ADIOI_BEEGFS_Sync_req_free -  */ int ADIOI_BEEGFS_Sync_req_free(void *extra_state) {     return MPI_SUCCESS; }

/*  * ADIOI_BEEGFS_Sync_req_cancel -  */ int ADIOI_BEEGFS_Sync_req_cancel(void *extra_state, int complete) {     return MPI_SUCCESS; }  

I have had a look at the implementation for MPI_File_iwrite() in ad_iwrite.c but this uses POSIX AIO, thus I am not sure I am doing things properly here.

the poll_fn() should be non-blocking. think MPI_Test more than unix poll/select/epoll. it's going to get called in a tight loop. Can you provide a stack trace for some of the stuck processes?

Additionally the problem does not show for small files (e.g. 1GB), which makes it not easy to debug.

BTW, I am using weak scaling thus every process always writes the same amount of data (64MB), I just change the number of procs in the test.

To test the modification with coll_perf I am using a configuration with 512 procs (64 nodes, 8procs/node) and 16 procs (8 nodes, 2procs/node). 

The 16 procs configuration writes 4 files of 1GB (16 x 64MB), the 512 procs configuration writes 4 files of 32GB (512 x 64MB).

Can someone spot any problem in my code?

Might take a few rounds but I'm sure we'll figure it out! ==rob

Thanks,

--  Giuseppe Congiu · Research Engineer II Seagate Technology, LLC office: +44 (0)23 9249 6082 · mobile:  www.seagate.com _______________________________________________ To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/devel