I am currently working on ROMIO developing a modification that enables the usage of locally attached SSDs as persistent cache in a HPC cluster. This extension is supposed to improve collective write operations performance, making them complete faster on local storage devices (scaling linearly with the number of aggregator nodes) and afterwards asynchronously flushing the local data to global file system while the application can progress with computation.
The extension uses the MPI Generalized request interface to provide non-blocking flushing of local data to global file system. I have added an ADIO_WriteContigLocal() function to the "common" ADIO driver that can be triggered by an apposite MPI hint (e.g. local_cache = enable) set by the application. The flushing of local data is started immediately after ADIOI_GEN_WriteStridedColl() returns to MPI_File_write_all(). The non-blocking syncing function starts a new pthread (I know this is not portable but the cluster I am using supports pthreads :-) ) and assigns it a file domain, which will be then read from the local file and written to the global file system.
When the file is finally closed, the implementation (still accordingly to the specific MPI hint) invokes an ADIO_CloseLocal() function which MPI_Wait(s) for all the pending generalized requests that have been started earlier. Finally, in ADIO_Close() I have a MPI_Barrier() to make sure that when MPI_File_close() returns all the file domains are consistent with the file in the global file system.
I am testing my code using IOR with the following configuration on a single node with 8 cores:
Strangely, when I increment the number of segments to be written by each process (i.e. -s) above 4/5 I get that IOR aborts with errors. Follows an example:
Command line used: ~/benchmarks/IOR/IOR -r -w -a MPIIO -b 1m -t 1m -c -s 5 -U ~/romio-totalview-dbg/ior.hints -o /work/ior11/testFile -H -V
Machine: Linux xxxx128
Summary:
api = MPIIO (version=3, subversion=0)
test filename = /work/ior11/testFile
access = single-shared-file
ordering in a file = sequential offsets
ordering inter file= no tasks offsets
clients = 8 (8 per node)
repetitions = 1
xfersize = 1 MiB
blocksize = 1 MiB
aggregate filesize = 40 MiB
hints passed to MPI_File_open() {
cb_buffer_size = 16777216
cb_nodes = 1
romio_cb_read = enable
romio_cb_write = enable
local_cache = enable
local_cache_path = /tmp/ior_tmp_file
local_cache_flush_flag = flush_immediate
local_cache_discard_flag = enable
romio_no_indep_rw = true
}
hints returned from opened file {
cb_buffer_size = 16777216
romio_cb_read = enable
romio_cb_write = enable
cb_nodes = 1
romio_no_indep_rw = true
romio_cb_pfr = disable
romio_cb_fr_types = aar
romio_cb_fr_alignment = 1
romio_cb_ds_threshold = 0
romio_cb_alltoall = automatic
ind_rd_buffer_size = 4194304
ind_wr_buffer_size = 524288
romio_ds_read = automatic
romio_ds_write = automatic
local_cache = enable
local_cache_flush_flag = flush_immediate
local_cache_discard_flag = enable
local_cache_path = /tmp/ior_tmp_file
cb_config_list = *:1
}
Assertion failed in file ~/mpich2/src/mpi/coll/helper_fns.c at line 491: status->MPI_TAG == recvtag
PSIlogger: Child with rank 1 exited with status 1.
(null)
Fatal error in PMPI_Barrier: Other MPI error, error stack:
PMPI_Barrier(428)......: MPI_Barrier(comm=0x84000001) failed
MPIR_Barrier_impl(335).: Failure during collective
MPIR_Barrier_impl(328).:
MPIR_Barrier(292)......:
MPIR_Barrier_intra(169):
mpid_irecv_done(101)...: read from socket failed - request state:recv(pde)done
Fatal error in PMPI_Barrier: Other MPI error, error stack:
PMPI_Barrier(428)......: MPI_Barrier(comm=0x84000001) failed
MPIR_Barrier_impl(335).: Failure during collective
MPIR_Barrier_impl(328).:
MPIR_Barrier(292)......:
MPIR_Barrier_intra(169):
PSIlogger: Child with rank 4 exited with status 1.
mpid_irecv_done(101)...: read from socket failed - request state:recv(pde)done
Fatal error in PMPI_Barrier: Other MPI error, error stack:
PMPI_Barrier(428)......: MPI_Barrier(comm=0x84000001) failed
MPIR_Barrier_impl(335).: Failure during collective
MPIR_Barrier_impl(328).:
MPIR_Barrier(292)......:
MPIR_Barrier_intra(169):
mpid_irecv_done(101)...: read from socket failed - request state:recv(pde)done
Fatal error in PMPI_Barrier: Other MPI error, error stack:
PMPI_Barrier(428)......: MPI_Barrier(comm=0x84000001) failed
MPIR_Barrier_impl(335).: Failure during collective
MPIR_Barrier_impl(328).:
MPIR_Barrier(292)......:
MPIR_Barrier_intra(169):
PSIlogger: Child with rank 3 exited with status 1.
mpid_irecv_done(101)...: read from socket failed - request state:recv(pde)done
Fatal error in PMPI_Barrier: Other MPI error, error stack:
PMPI_Barrier(428)......: MPI_Barrier(comm=0x84000000) failed
MPIR_Barrier_impl(335).: Failure during collective
MPIR_Barrier_impl(328).:
MPIR_Barrier(292)......:
MPIR_Barrier_intra(169):
mpid_irecv_done(101)...: read from socket failed - request state:recv(pde)done
MPIR_Barrier_intra(169):
PSIlogger: Child with rank 5 exited with status 1.
mpid_irecv_done(101)...: read from socket failed - request state:recv(pde)done
wait entry: 2a1ef18, 21f2790, 2a1e8e8, 21f4488
Fatal error in PMPI_Wait: Other MPI error, error stack:
PMPI_Wait(180)..........: MPI_Wait(request=0x21f4488, status=0x21f2138) failed
MPIR_Wait_impl(77)......:
MPIR_Grequest_query(447): user request query function returned error code 1601
Fatal error in PMPI_Barrier: Other MPI error, error stack:
PMPI_Barrier(428)......: MPI_Barrier(comm=0x84000000) failed
MPIR_Barrier_impl(335).: Failure during collective
MPIR_Barrier_impl(328).:
MPIR_Barrier(292)......:
MPIR_Barrier_intra(169):
mpid_irecv_done(101)...: read from socket failed - request state:recv(pde)done
MPIR_Barrier_intra(169):
mpid_irecv_done(101)...: read from socket failed - request state:recv(pde)done
MPIR_Barrier_intra(169):
PSIlogger: Child with rank 2 exited with status 1.
PSIlogger: Child with rank 7 exited with status 1.
PSIlogger: Child with rank 6 exited with status 1.
PSIlogger: Child with rank 0 exited with status 1.
mpid_irecv_done(101)...: read from socket failed - request state:recv(pde)done
To further provide some description of what is happening here: when I increment the number of segments, IOR will increment the number of calls to MPI_File_write_all() (made one after the other). For every collective write my extension keeps track of the new file domain and starts a new thread to "handle" it. Eventually, there might be a number of pthread (active at the same time) equal to the number of segments, which should not be a problem considering small numbers like 5 or 10.