Re: [squid-users] My my squid hit ratio so low?

On 5/07/2013 7:04 p.m., jinge wrote:
> Dear Amos,
>
> Thanks for your reply.
>
>
> On 2013-7-5, at 下午2:39, Amos Jeffries <squid3_at_treenet.co.nz
> <mailto:squid3_at_treenet.co.nz>> wrote:
>
>> On 5/07/2013 6:26 p.m., jinge wrote:
>>> Hi, all.
>>>
>>> We use squid for a long time. And recently we upgrade our squid to
>>> 3.3.4 and begin to use SMP and rock.
>>>
>>> This is our related configure.
>>>
>>> include /usr/local/etc/squid/commonbk/bk.global-options.conf
>>> include /usr/local/etc/squid/commonbk/bk.refresh-pattern.conf
>>> include /usr/local/etc/squid/commonbk/bk.acl-define.conf
>>> include /usr/local/etc/squid/commonbk/bk.acl-action.conf
>>> #include /usr/local/etc/squid/squid.conf
>>> cache_dir rock /cache1/rock 48000 max-size=31000 max-swap-rate=300
>>> swap-timeout=300
>>> cache_dir rock /cache2/rock 48000 max-size=31000 max-swap-rate=300
>>> swap-timeout=300
>>> workers 3
>>> cpu_affinity_map process_numbers=1,2,3 cores=3,5,7
>>> if ${process_number} = 1
>>> include /usr/local/etc/squid/commonbk/backend5a.conf
>>> endif
>>> if ${process_number} = 2
>>> include /usr/local/etc/squid/commonbk/backend5b.conf
>>> endif
>>> if ${process_number} = 3
>>> include /usr/local/etc/squid/commonbk/backend5c.conf
>>> endif
>>
>> Lot of sub-config files there. What do they contain?
>
> And the sub-config is some thing like this.
> # NETWORK
> #
> ---------------------------------------------------------------------------
> http_port 192.168.2.1:3128 accel allow-direct ignore-cc
> # DISK
> #
> -----------------------------------------------------------------------------
> cache_dir diskd /cache3/aufs/64k 24000 16 128 min-size=31001
> max-size=65536
>

So 64KB maximum cacheable object? at least for that worker process.

>>
>>> access_log /dev/null
>>
>> Do not waste rsources formatting log lines only to send them to
>> /dev/null.
>>
>> "access_log none" does what you want in a far more efficient way.
>
> Thank you, I will follow your advise.
>
>>
>>> cache_log /var/log/squid/cache.log
>>>
>>>
>>> and our machine
>>>
>>> No LSB modules are available.
>>> Distributor ID:Ubuntu
>>> Description:Ubuntu Raring Ringtail (development branch)
>>> Release:13.04
>>> Codename:raring
>>>
>>> total used free shared buffers
>>> cached
>>> Mem: 15G 14G 1.5G 0B 27M
>>> 6.7G
>>> -/+ buffers/cache: 7.5G 8.1G
>>> Swap: 16G 0B 16G
>>>
>>> Linux 3.8.0-16-generic #26-Ubuntu SMP Mon Apr 1 19:52:57 UTC 2013
>>> x86_64 x86_64 x86_64 GNU/Linux
>>>
>>> /dev/sdb1 137G 13G 125G 9% /cache1
>>> /dev/sdc1 69G 13G 57G 18% /cache2
>>> /dev/sdd1 69G 37G 33G 54% /cache3
>>> /dev/sde1 69G 31G 38G 45% /cache4
>>> /dev/sdf1 69G 31G 39G 45% /cache5
>>>
>>> we found that the Rock won't fill the cache_dir in 48GB. And our
>>> ratio is so low
>>>
>>> Hits as % of all requests:5min: 4.3%, 60min: 4.3%
>>> Hits as % of bytes sent:5min: 3.3%, 60min: 3.3%
>>> Memory hits as % of hit requests:5min: 33.7%, 60min: 32.1%
>>> Disk hits as % of hit requests:5min: 8.5%, 60min: 9.0%

These numbers are not adding up at all. For example 32% + 9% is not 100%
of HIT requests - it is ~50% of requests (2 workers?).
Possibly your real numbers are twice those % points. Still low though.
Should be up around 80-90% for a reverse-proxy like you show in the
config above.

>>>
>>>
>>> Can anyone tell me what's wrong with my squid?
>>
>> Not from just those numbers. This is where an access.log comes in
>> handy. For identifying where the MISSes are and if any should have
>> been HITs.
>>
>>
>
> And to my surprise, we after a long time of running. Why my rock cache
> dir won't fill the whole storage as I config:
> cache_dir rock /cache1/rock 48000 max-size=31000 max-swap-rate=300
> swap-timeout=300
>
> Does the 4800 directive won't work, or there is any other reason?

That depends on what your website(s) contains in the way of objects.
Very likely you have not served up 48GB worth of unique under-30.274 KB
sized objects.

If you are able to I suggest a little experiment. Enable your access.log
again and gather some statistics about the following:
a) what count of objects are flowing through your Squid are under 31000
bytes and HIT vs MISS?

   This will tell you what your Rock cache HIT-rate is. IME we should
expect something roughly similar for the over-31000 byte objects as well
although there may be less of them contributing towards the total HIT rates.

b) what count of objects are flowing through your Squid are over 31000
bytes and HIT vs MISS?

This will give you an idea of how much the UFS cache_dir are warping the
overall ratios. Not being SMP-aware a request cached by one worker
cannot be HIT on another, so yoru ration may be up to 60% decreased with
3 workers if the traffic is mostly going via these caches.

c) what count of objects are flowing through your Squid are over 64KB
and any other cache_dir max-size you have configured in the per-worker
configs?

This will give you an idea of whether those max-size limits are
reasonable. For example if that worker whose sub-config you showed with
cache3 max-size of 64KB was mostly receiving 70KB objects - it would
have a very poor HIT ratio just because of that limit alone.

d) what the request method and response codes are? (both HTTP satus code
and Squid log code).

This will give you a clue whether it is the nature of teh traffic itself
(revalidation and/or nasty server response to revalidation, or
non-cacheable methods being used by clients) is causing low HIT ratios.

Amos
Received on Fri Jul 05 2013 - 08:31:47 MDT