Good catch :) Looks correct to me.
From: 池信泽 [mailto:[email protected]]
Sent: Tuesday, October 28, 2014 9:24 AM
To: Sage Weil
Cc: Wang, Zhiqiang; [email protected]; [email protected]
Subject: Re: Cache pool bug?
I think there is also bug in ReplicatedPG::agent_estimate_atime_temp
I think we should change the following code:
for (map<time_t,HitSetRef>::iterator p = agent_state->hit_set_map.begin();
p != agent_state->hit_set_map.end();
++p) {
if (p->second->contains(oid)) {
if (*atime < 0)
*atime = now - p->first;
if (temp)
++(*temp);
else
return;
}
}
to something like this:
for (map<time_t,HitSetRef>::iterator p = agent_state->hit_set_map.rbegin();
p != agent_state->hit_set_map.rend();
++p) {
if (p->second->contains(oid)) {
if (*atime < 0)
*atime = now - p->first;
if (temp)
++(*temp);
else
return;
}
}
Because if there are mutiple time_t in agent_state for the same object, we
should use the recently one.
2014-10-28 0:04 GMT+08:00 Sage Weil <[email protected]>:
On Mon, 27 Oct 2014, Wang, Zhiqiang wrote:
> Hi,
>
> I agree with you this is a bug.
>
> I think we should change the following code:
>
> // KISS: if [lower,upper] spans our target effort, evict it.
> if (atime_lower >= agent_state->evict_effort)
> return false;
>
> to something like this:
>
> if (1000000 - atime_upper >= agent_state->evict_effort)
> return false;
>
> cc the ceph-devel list for comments.
Yes, that does look backwards. Good catch! If you submit a pull request
we can get it merged and then backported.
Thanks!
sage
> ________________________________________
> Date: Mon, 27 Oct 2014 10:20:01 +0800
> Subject: Cache pool bug?
> From: [email protected]
> To: [email protected]
> hi, zhiqiang?
>
> When I read the ReplicatedPG::agent_maybe_evic, I found than the newest
> object has
>
> more probability to be evicted. For Example:
>
> the object in hit_set is newest object, call
>
> the funtion agent_estimate_atime_temp(soid, &atime, NULL /*FIXME
> &temp*/); atime is 0 when return.
>
> agent_state->atime_hist.get_position_micro(atime, &atime_lower,
> &atime_upper), the atime_lower is 0 when return.
>
> // KISS: if [lower,upper] spans our target effort, evict it.
> if (atime_lower >= agent_state->evict_effort)
> return false;
>
> Above, the newest object in hit_set is evicted.
>
> Is there a bug? Or, whether I skip something?
>
>
> Thanks
> N?????r??y??????X???v???)?{.n?????z?]z????ay? ????j ??f???h????? ?w???
???j:+v???w???????? ????zZ+???????j"????i
