Getting the time when each batch of pages starts is
sufficiently accurate, and ktime_get() is several % of the
CPU time when doing AIO + DIO.
This relies on previous patches in this series.
Measuring this in milliseconds/gigabyte lets us measure the
improvement in absolute terms, rather than just relative
terms.
This patch reduces i/o time in ms/GiB by:
Write: 17 ms/GiB
Read: 6 ms/GiB
Totals:
Write: 237 ms/GiB
Read: 223 ms/GiB
IOR:
mpirun -np 1 $IOR -w -r -t 64M -b 64G -o ./iorfile --posix.odirect
Without the patch:
write 4030 MiB/s
read 4468 MiB/s
With patch:
write 4326 MiB/s
read 4587 MiB/s
Signed-off-by: Patrick Farrell <farr0186@gmail.com>
Change-Id: I02897bf810683bc77a7d09156cdb83ba1d25ebf1
Reviewed-on: https://review.whamcloud.com/39437
Reviewed-by: Wang Shilong <wshilong@whamcloud.com>
Tested-by: jenkins <devops@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
pgoff_t start, pgoff_t end);
void osc_lru_add_batch(struct client_obd *cli, struct list_head *list);
void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
pgoff_t start, pgoff_t end);
void osc_lru_add_batch(struct client_obd *cli, struct list_head *list);
void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
- enum cl_req_type crt, int brw_flags);
+ enum cl_req_type crt, int brw_flags, ktime_t submit_time);
int lru_queue_work(const struct lu_env *env, void *data);
long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
long target, bool force);
int lru_queue_work(const struct lu_env *env, void *data);
long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
long target, bool force);
unsigned int max_pages;
unsigned int ppc_bits; /* pages per chunk bits */
unsigned int ppc;
unsigned int max_pages;
unsigned int ppc_bits; /* pages per chunk bits */
unsigned int ppc;
+ ktime_t submit_time = ktime_get();
bool sync_queue = false;
LASSERT(qin->pl_nr > 0);
bool sync_queue = false;
LASSERT(qin->pl_nr > 0);
oap->oap_async_flags |= ASYNC_COUNT_STABLE;
spin_unlock(&oap->oap_lock);
oap->oap_async_flags |= ASYNC_COUNT_STABLE;
spin_unlock(&oap->oap_lock);
- osc_page_submit(env, opg, crt, brw_flags);
+ osc_page_submit(env, opg, crt, brw_flags, submit_time);
list_add_tail(&oap->oap_pending_item, &list);
if (page->cp_sync_io != NULL)
list_add_tail(&oap->oap_pending_item, &list);
if (page->cp_sync_io != NULL)
* transfer (i.e., transferred synchronously).
*/
void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
* transfer (i.e., transferred synchronously).
*/
void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
- enum cl_req_type crt, int brw_flags)
+ enum cl_req_type crt, int brw_flags, ktime_t submit_time)
{
struct osc_io *oio = osc_env_io(env);
struct osc_async_page *oap = &opg->ops_oap;
{
struct osc_io *oio = osc_env_io(env);
struct osc_async_page *oap = &opg->ops_oap;
oap->oap_cmd |= OBD_BRW_NOQUOTA;
}
oap->oap_cmd |= OBD_BRW_NOQUOTA;
}
- opg->ops_submit_time = ktime_get();
+ opg->ops_submit_time = submit_time;
osc_page_transfer_get(opg, "transfer\0imm");
osc_page_transfer_add(env, opg, crt);
}
osc_page_transfer_get(opg, "transfer\0imm");
osc_page_transfer_add(env, opg, crt);
}