limit writes to backend device for reads

[disaster123]

bcache has a pretty cool function if the drive can't keep up with the writes or is too slow for the writes it adds a dynamic sleep between each backend write.

Currently i've the problem that latency for reads is very high on dm-writeboost devices if you have a lot of writes.

Stefan

[akiradeveloper]

hi

writeback_threshold doesnt work? I think that's what you want

Sent from my iPhone

On 2015/07/23, at 19:44, disaster123 notifications@github.com wrote:

bcache has a pretty cool function if the drive can't keep up with the writes or is too slow for the writes it adds a dynamic sleep between each backend write.

Currently i've the problem that latency for reads is very high on dm-writeboost devices if you have a lot of writes.

Stefan

― Reply to this email directly or view it on GitHub.

[disaster123]

ah OK i'm sorry i think i've misread the documentaion reading the code makes it more clean:

int writeback_modulator_proc(void *data)
{
        struct wb_device *wb = data;

        struct hd_struct *hd = wb->backing_dev->bdev->bd_part;
        unsigned long old = 0, new, util;
        unsigned long intvl = 1000;

        while (!kthread_should_stop()) {
                new = jiffies_to_msecs(part_stat_read(hd, io_ticks));

                util = div_u64(100 * (new - old), 1000);

                if (util < ACCESS_ONCE(wb->writeback_threshold))
                        wb->allow_writeback = true;
                else
                        wb->allow_writeback = false;

                old = new;

                schedule_timeout_interruptible(msecs_to_jiffies(intvl));
        }
        return 0;
}
'''

Will redo some tests with lowering the value to 50.

[disaster123]

mhm but isn't that too easy? It is OK to have 100% writes if there are no reads.

[akiradeveloper]

Will redo some tests with lowering the value to 50.

The calculation of load is approximation. or there could be a mistake. But it actually turna off writeback when the load is too high. That's enough for me.

It is OK to have 100% writes if there are no reads.

I don't think so. Keeping the load around 70 eventually achieves the highest throughput.

[disaster123]

Is it the same algorithm iostat is using?

[akiradeveloper]

I believe yes but I am not sure about iostat other than it looks at diskstats.

I learned from the diskstats the correct usage of part_stat_read.

http://lxr.free-electrons.com/source/block/genhd.c#L1147

1163         disk_part_iter_init(&piter, gp, DISK_PITER_INCL_EMPTY_PART0);
1164         while ((hd = disk_part_iter_next(&piter))) {
1165                 cpu = part_stat_lock();
1166                 part_round_stats(cpu, hd);
1167                 part_stat_unlock();
1168                 seq_printf(seqf, "%4d %7d %s %lu %lu %lu "
1169                            "%u %lu %lu %lu %u %u %u %u\n",
1170                            MAJOR(part_devt(hd)), MINOR(part_devt(hd)),
1171                            disk_name(gp, hd->partno, buf),
1172                            part_stat_read(hd, ios[READ]),
1173                            part_stat_read(hd, merges[READ]),
1174                            part_stat_read(hd, sectors[READ]),
1175                            jiffies_to_msecs(part_stat_read(hd, ticks[READ])),
1176                            part_stat_read(hd, ios[WRITE]),
1177                            part_stat_read(hd, merges[WRITE]),
1178                            part_stat_read(hd, sectors[WRITE]),
1179                            jiffies_to_msecs(part_stat_read(hd, ticks[WRITE])),
1180                            part_in_flight(hd),
1181                            jiffies_to_msecs(part_stat_read(hd, io_ticks)),
1182                            jiffies_to_msecs(part_stat_read(hd, time_in_queue))