Re: Terrible performance of sequential O_DIRECT 4k writes in SANenvironment. ~3 times slower then Solars 10 with the same HBA/Storage.

From: Sergey Meirovich
Date: Fri Jan 10 2014 - 09:33:01 EST


Hi Jan,

On 10 January 2014 12:48, Jan Kara <jack@xxxxxxx> wrote:
> On Fri 10-01-14 12:36:22, Sergey Meirovich wrote:
>> Hi Jan,
>>
>> On 10 January 2014 11:36, Jan Kara <jack@xxxxxxx> wrote:
>> > On Thu 09-01-14 12:11:16, Sergey Meirovich wrote:
>> ...
>> >> I've done preallocation on fnic/XtremIO as Christoph suggested.
>> >>
>> >> [root@dca-poc-gtsxdb3 mnt]# sysbench --max-requests=0
>> >> --file-extra-flags=direct --test=fileio --num-threads=4
>> >> --file-total-size=10G --file-io-mode=async --file-async-backlog=1024
>> >> --file-rw-ratio=1 --file-fsync-freq=0 --max-requests=0
>> >> --file-test-mode=seqwr --max-time=100 --file-block-size=4K prepare
>> >> sysbench 0.4.12: multi-threaded system evaluation benchmark
>> >>
>> >> 128 files, 81920Kb each, 10240Mb total
>> >> Creating files for the test...
>> >> [root@dca-poc-gtsxdb3 mnt]# du -k test_file.* | awk '{print $1}' |sort |uniq
>> >> 81920
>> >> [root@dca-poc-gtsxdb3 mnt]# fallocate -l 81920k test_file.*
>> >>
>> >> Results: 13.042Mb/sec 3338.73 Requests/sec
>> >>
>> >> Probably sysbench is still triggering append DIO scenario. Will say
>> >> simple wrapper over io_submit() against already preallocated (and even
>> >> filled with data) file provide much better throughput if your theory
>> >> is valid?
>> > So I was experimenting a bit. "sysbench prepare" seems to always do
>> > synchronous IO from a single thread in the 'prepare' phase regardless of
>> > the arguments. So there the reported throughput isn't really relevant.
>> >
>> > In the 'run' phase it obeys the arguments and indeed when I run fallocate
>> > to preallocate files during 'run' phase, it significantly helps the
>> > throughput (from 20 MB/s to 55 MB/s on my SATA drive).
>>
>> Sorry, Jan. Seems that I presented my findings in a previous mail in
>> ambiguous style . I know that prepare phase of sysbench is
>> synchronous/probably buffered (because I saw 512k chunks sent down to
>> HBA)? IO. I played with blocktrace and have seen that myself during
>> prepare:
>>
>> [root@dca-poc-gtsxdb3 mnt]# sysbench --max-requests=0
>> --file-extra-flags=direct --test=fileio --num-threads=4
>> --file-total-size=10G --file-io-mode=async --file-async-backlog=1024
>> --file-rw-ratio=1 --file-fsync-freq=0 --max-requests=0
>> --file-test-mode=seqwr --max-time=100 --file-block-size=4K prepare
>> ...
>>
>> Leads to:
>>
>> [root@dca-poc-gtsxdb3 mnt]# blktrace -d /dev/sdg -o - | blkparse -i -
>> | grep 'D W'
>> 8,96 14 604 53.129805520 28114 D WS 1116160 + 1024 [sysbench]
>> 8,96 14 607 53.129843345 28114 D WS 1120256 + 1024 [sysbench]
>> 8,96 14 610 53.129873782 28114 D WS 1124352 + 1024 [sysbench]
>> 8,96 14 613 53.129903703 28114 D WS 1128448 + 1024 [sysbench]
>> 8,96 14 616 53.130957213 28114 D WS 1132544 + 1024 [sysbench]
>> 8,96 14 619 53.130988835 28114 D WS 1136640 + 1024 [sysbench]
>> 8,96 14 622 53.131018854 28114 D WS 1140736 + 1024 [sysbench]
>> ...
> Ah, ok. I misuderstood what you wrote then.
>
>> That result "13.042Mb/sec 3338.73 Requests/sec" was from run phase
>> and before it fallocate had been made.
>>
>> blktrace from run phase looks very different. 4k as expected.
>> [root@dca-poc-gtsxdb3 ~]# blktrace -d /dev/sdg -o - | blkparse -i - |
>> grep 'D W'
>> 8,96 5 3 0.000001874 28212 D WS 1847296 + 8 [sysbench]
>> 8,96 5 7 0.001213728 28212 D WS 1847304 + 8 [sysbench]
>> 8,96 5 11 0.002779304 28212 D WS 1847312 + 8 [sysbench]
>> 8,96 5 15 0.004486445 28212 D WS 1847320 + 8 [sysbench]
>> 8,96 5 19 0.006012133 28212 D WS 22691864 + 8 [sysbench]
>> 8,96 5 23 0.007781553 28212 D WS 22691896 + 8 [sysbench]
>> 8,96 5 27 0.009043404 28212 D WS 22691928 + 8 [sysbench]
>> 8,96 5 31 0.010546829 28212 D WS 22691960 + 8 [sysbench]
>> 8,96 5 35 0.012214468 28212 D WS 22691992 + 8 [sysbench]
>> 8,96 5 39 0.013792616 28212 D WS 22692024 + 8 [sysbench]
>> ...
> Strange - I see:
> 8,32 7 2 0.000086080 0 D WS 1869752 + 1024 [swapper]
> 8,32 7 7 0.041126425 0 D WS 1874712 + 24 [swapper]
> 8,32 7 6 0.041054543 0 D WS 1871792 + 416 [swapper]
> 8,32 7 7 0.041126425 0 D WS 1874712 + 24 [swapper]
> 8,32 6 118 0.042761949 28952 D WS 1875416 + 528 [sysbench]
> 8,32 6 143 0.042995928 28952 D WS 1876888 + 48 [sysbench]
> 8,32 5 352 0.045154160 28955 D WS 1876936 + 168 [sysbench]
> 8,32 6 444 0.045527660 28952 D WS 1878296 + 992 [sysbench]
> ...
>
> Not ideal but significantly better. The only idea I have: Didn't you run
> fallocate(1) before you started the 'run' phase? Because 'run' phase
> truncates the files before doing io to them. Can you check that during run
> phase (after fallocate is run) the file size is constantly at 80MB?

Jan, I believe your initial theory that append AIO is equal to
Synchronous DIO is absolutely correct. I've given up sysbench and
written simple dirty wrapper around io_submit() and run it against
preallocated file.. XtremIO is doing online deduplication so results
were wonderful

694.25 MB/s 177728.84 Req/sec

[root@dca-poc-gtsxdb3 mnt]# dd if=/dev/zero of=4k.data bs=4096 count=524288
524288+0 records in
524288+0 records out
2147483648 bytes (2.1 GB) copied, 5.75357 s, 373 MB/s
[root@dca-poc-gtsxdb3 mnt]# ./4k
io_submit() accepted 524288 IOs
io_getevents() returned 524288 events
time elapsed (sec.): 2.949932
bandwidth (MiB/s): 694.25
IOps: 177728.84
[root@dca-poc-gtsxdb3 mnt]#
...
[root@dca-poc-gtsxdb3 ~]# vmstat 1
procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu-----
r b swpd free buff cache si so bi bo in cs us sy id wa st
0 0 0 260598160 260540 1896032 0 0 3 24 11 12
0 0 100 0 0
0 0 0 260598256 260544 1896032 0 0 0 16 1106 378
0 0 100 0 0
0 0 0 260598288 260544 1896032 0 0 0 0 1093 373
0 0 100 0 0
1 0 0 260499536 260544 1928368 0 0 0 293820 2782 1438
0 1 99 0 0
2 0 0 260484816 260544 1928804 0 0 0 820152 5575 3146
0 3 97 0 0
1 0 0 260481680 260544 1928804 0 0 0 710028 4844 2947
0 3 97 0 0
1 0 0 260548384 260544 1928804 0 0 0 273168 2187 1744
0 2 98 0 0
0 0 0 260549088 260544 1928804 0 0 0 4 1156 426
0 0 100 0 0
0 0 0 260549472 260544 1928804 0 0 0 0 1082 328
0 0 100 0 0
^C
[root@dca-poc-gtsxdb3 ~]#


========================== io_submit() wrapper =============================
#define _GNU_SOURCE

#include <errno.h>
#include <libaio.h>

#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>

#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#include <unistd.h>
#include <sys/time.h>


#define FNAME "4k.data"
#define IOSIZE 4096
#define REQUESTS 524288

/* gcc 4k.c -std=gnu99 -laio -o 4k */

int main(void) {
io_context_t ctx;
int ret;

int flag = O_RDWR | O_DIRECT;
int fd = open(FNAME, flag);
if (fd == -1) {
printf("open(%s, %d) - failed!\nExiting.\n"
"If file doesn't exist please precreate it "
"with dd if=/dev/zero of=%s bs=%d count=%d\n",
FNAME, flag, FNAME, IOSIZE, REQUESTS);
return errno;
}

memset(&ctx, 0, sizeof(io_context_t));
if (io_setup(REQUESTS, &ctx)) {
printf("io_setup(%d, &ctx) failed\n", REQUESTS);
return -ret;
}

void *mem = NULL;
posix_memalign(&mem, 4096, IOSIZE);
memset(mem, 9, IOSIZE);
struct iocb *aio = malloc(sizeof(struct iocb) * REQUESTS);
memset(aio, 0, sizeof(struct iocb) * REQUESTS);
struct iocb **lio = malloc(sizeof(void *) * REQUESTS);
memset(lio, 0, sizeof(void *) * REQUESTS);
struct io_event *event = malloc(sizeof(struct io_event) * REQUESTS);
memset(event, 0, sizeof(struct io_event) * REQUESTS);

for (int i = 0; i < REQUESTS; i++) {
io_prep_pwrite(&aio[i], fd, mem, IOSIZE, i * IOSIZE);
lio[i] = &aio[i];
}
struct timeval start, end;
gettimeofday(&start, NULL);
ret = io_submit(ctx, REQUESTS, lio);
printf("io_submit() accepted %d IOs\n", ret);
fdatasync(fd);

ret = io_getevents(ctx, REQUESTS, REQUESTS, event, NULL);
printf("io_getevents() returned %d events\n", ret);
gettimeofday(&end, NULL);

double elapsed = (end.tv_sec - start.tv_sec) +
((end.tv_usec - start.tv_usec)/1000000.0);
printf("time elapsed (sec.):\t%2f\n", elapsed);
printf("bandwidth (MiB/s):\t%.2f\n",
(double) (((long long) IOSIZE * REQUESTS) / (1024 * 1024))
/ elapsed);
printf("IOps:\t\t\t%.2f\n", (double) REQUESTS
/ elapsed);

if (io_destroy(ctx)) {
perror("io_destroy");
return -1;
}

free(aio);
free(lio);
free(event);

return 0;
}


>
> Honza
> --
> Jan Kara <jack@xxxxxxx>
> SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/