On 2012-07-14 07:56 Dan Williams [off-list ref] Wrote:
[ adding Shaohua ]
On Fri, Jul 13, 2012 at 3:31 AM, majianpeng [off-list ref] wrote:
quoted
To improve write perfomance by decreasing the preread stripe,only move
IO_THRESHOLD stripes from delay_list to hold_list once.
Using the follow command:
dd if=/dev/zero of=/dev/md0 bs=2M count=52100.
At default condition: speed is 95MB/s.
At the condition of preread_bypass_threshold was equal zero:speed is 105MB/s.
Using this patch:speed is 123MB/s.
If preread_bypass_threshold was zero,the performance will be better,but
not better than this patch.
I think maybe two reason:
1:If bio is REQ_SYNC
2:In function __get_priority_stripe():
quoted
quoted
} else if (!list_empty(&conf->hold_list) &&
((conf->bypass_threshold &&
conf->bypass_count > conf->bypass_threshold) ||
atomic_read(&conf->pending_full_writes) == 0)) {
Preread_bypass_threshold is one condition of getting stripe from
hold_list.So only control the number of hold_list can get better
performance.
So this is a pretty obvious tradeoff of increased latency for improved
throughput. Any idea how much this change affects latency?
Especially in the fast device case?
I did not think the latency.If it only fetch preread_bypass_threshold stripes from delay_list to
host_list,the latency can be control by userspace.
The code like :
static void raid5_activate_delayed(struct r5conf *conf)
{
+ int count = 0;
if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
while (!list_empty(&conf->delayed_list)) {
struct list_head *l = conf->delayed_list.next;@@ -3672,6 +3673,8 @@ static void raid5_activate_delayed(struct r5conf *conf)
if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
atomic_inc(&conf->preread_active_stripes);
list_add_tail(&sh->lru, &conf->hold_list);
+ if (++count >= conf->preread_active_stripes)
+ break;
}