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: xmdxcxz@xxxxxxxxx
> To: wonzhq@xxxxxxxxxxx
> 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
