Testing has shown the ll_inode_inode's lli_write_mutex to be a
limiting factor with single shared file write performance, when using
many writing threads on a single machine. Even if each thread is
writing to a unique portion of the file, the lli_write_mutex will
prevent no more than a single thread to ever write to the file
simultaneously.

This change attempts to remove this bottle neck, by replacing this
mutex with a range lock. This should allow multiple threads to write
to a single file simultaneously iff the threads are writing to unique
regions of the file.

Performance testing shows this change to garner a significant
performance boost to write bandwidth. Using FIO on a single machine
with 32 GB of RAM, write performance tests were run with and without
this change applied; the results are below:

    +---------+-----------+---------+--------+--------+--------+
    |     fio v2.0.13     |        Write Bandwidth (KB/s)      |
    +---------+-----------+---------+--------+--------+--------+
    | # Tasks | GB / Task | Test 1  | Test 2 | Test 3 | Test 4 |
    +---------+-----------+---------+--------+--------+--------+
    |    1    |    64     |  452446 | 454623 | 457653 | 463737 |
    |    2    |    32     |  850318 | 565373 | 602498 | 733027 |
    |    4    |    16     | 1058900 | 463546 | 529107 | 976284 |
    |    8    |     8     | 1026300 | 468190 | 576451 | 963404 |
    |   16    |     4     | 1065500 | 503160 | 462902 | 830065 |
    |   32    |     2     | 1068600 | 462228 | 466963 | 749733 |
    |   64    |     1     |  991830 | 556618 | 557863 | 710912 |
    +---------+-----------+---------+--------+--------+--------+

 * Test 1: Lustre client running 04ec54f. File per process write
           workload. This test was used as a baseline for what we
           _could_ achieve in the single shared file tests if the
           bottle necks were removed.

 * Test 2: Lustre client running 04ec54f. Single shared file
           workload, each task writing to a unique region.

 * Test 3: Lustre client running 04ec54f + I0023132b. Single shared
           file workload, each task writing to a unique region.

 * Test 4: Lustre client running 04ec54f + this patch.
           Single shared file workload, each task writing to a unique
           region.

Direct IO does not use the page cache like normal IO, so
concurrent direct IO reads of the same pages are not safe.

As a result, direct IO reads must take the range lock
in ll_file_io_generic, otherwise they will attempt to work
on the same pages and hit assertions like:
(osc_request.c:1219:osc_brw_prep_request())
ASSERTION( i == 0 || pg->off > pg_prev->off ) failed:
i 3 p_c 10 pg ffffea00017a5208 [pri 0 ind 2771] off 16384
prev_pg ffffea00017a51d0 [pri 0 ind 2256] off 16384

Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Patrick Farrell <paf@cray.com>
Signed-off-by: Jinshan Xiong <jinshan.xiong@intel.com>
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-1669
Reviewed-on: http://review.whamcloud.com/6320
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6227
Reviewed-on: http://review.whamcloud.com/14385
Reviewed-by: Bobi Jam <bobijam@hotmail.com>
Reviewed-by: Alexander Boyko <alexander.boyko@seagate.com>
Reviewed-by: Hiroya Nozaki <nozaki.hiroya@jp.fujitsu.com>
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>
Signed-off-by: James Simmons <jsimmons@infradead.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

0 files changed, 0 insertions, 0 deletions