[squid-users] AUFS vs. DISKS

[Marcus Kool]

On 07/15/2015 11:39 AM, Amos Jeffries wrote:
> On 16/07/2015 1:51 a.m., Stakres wrote:
>> Hi Fred,
>> tests from my side:
>> DISKD with TCP_HIT objects: 564KB/s with wget, the same url you have tested.
>> AUFS with TCP_HITS objects: 47.8M/s, same wget, same squid, same url, same
>> all.
>>
>> Wget with AUFS:
>> Length: 10095849 (9.6M) [application/x-msdos-program]
>> Saving to: `youtube_downloader_hd_setup-2.9.9.23.exe'
>> 100%[======================================>] 10,095,849  47.9M/s   in 0.2s
>> 2015-07-15 15:48:29 (47.9 MB/s) - `youtube_downloader_hd_setup-2.9.9.23.exe'
>> saved
>>
>> All,
>> We have switched some ISPs from DISKD to AUFS this morning, the "queue
>> congestion" appears at the begining then disappears from the cache.log. For
>> how long, nobody knows...
>
>
> Doh. Sorry its been so long I had forgotten one detail ...
>
> Squid AUFS starts with a relatively short I/O queue length (8 entries)
> and auto-grows it as needed by doubling the limit and displaying the
> congestion message. On any busy proxy this can be seen for a while after
> startup, then possibly during the first traffic peak after that. It is
> not a problem unless it keeps growing indefinitely or appears suddenly
> without a matching traffic spike (possible sign of disk failure).
>
> I'm applying a patch to raise the baseline queue length from 8 to 8K,
> and reporting in each message what the limit was changed to. That should
> get rid of a dozen entries in your logs, and help show whats going on in
> the remainder.

Since you will modify the log message from
    WARNING - Queue congestion
to something like
    INFO - the I/O queue lenght was X and has been doubled to 2X because of the current I/O load.
admins will understand _much_ better what is going on.

I think that changing the baseline to 8K is not required since the queue congestion
warning is normally seen only a few times, so the baseline value of 8 is doubled
only a few times.
A new baseline value of 256 (5 doublings) makes sense to me since most disks have
a hardware queue depth of 256.
For those systems where many disks are tortured with I/O the doubling will appear
a few (log2(#disks)) times.

Marcus

> Amos
>
> _______________________________________________
> squid-users mailing list
> squid-users at lists.squid-cache.org
> http://lists.squid-cache.org/listinfo/squid-users
>