Re: [squid-users] Object Hit/Byte Hit accounting with Multiple Instances

On 16/12/2010, at 12:44 PM, Amos Jeffries wrote:
> On 15/12/10 14:38, Michael Hendrie wrote:
>> Hello List,
>>
>> I have server running 3 instances of squid-3.0.STABLE19 using a
>> configuration similar to that documented at
>> http://wiki.squid-cache.org/MultipleInstances. Each instance has all
>> other instance configured as siblings using the "proxy-only"
>> directive
>> to allow sharing of cache without duplicating objects. This setup is
>> working very well and has increased server performance by over 50%.
>>
>> I'm now trying to get an accurate indication of byte savings I'm
>> achieving with this configuration however I'm not sure that the
>> calculations I'm using are giving the correct results. Because each
>> instance maintains a separate cache_dir this seems to be a little
>> difficult to calculate. When instance 1 records a request as a MISS
>> it
>> may in fact be a HIT (from an entire system point of view) if the
>> object
>> is retrieved from the cache of instance 2 or 3.
>>
>> Using a combination of "squidclient mgr:counters" and SNMP, I grab
>> counter values from each instance, tally and use the following
>> formula
>> to calculate the byte hit ratio:
>>
>> (mgr:counters:client_http.hit_kbytes_out +
>> snmp:cacheClientHTTPHitKb.sibling_addresses) /
>> (mgr:counters:client_http.kbytes_out -
>> snmp:cacheClientHTTPHitKb.sibling_addresses) * 100 = % cache byte
>> hit ratio
>>
>> Using this formula, I always seem to get inconsistencies between what
>> squid reports and what my benchmarking tool reports (web-
>> polygraph). In
>> the few cases I've checked so far, squid is always reporting a 4-5%
>> less
>> byte hit than what web-polygraph reports.
>
> That sounds about the size of header overheads to me.
> Give 3.2 workers a try out now and see if that is usable. The stats
> calculations are fixed there for multiple workers.
>

Unfortunately I must use this version (for the moment) for reasons
beyond my control. Just to clarify

1). Are you saying that headers aren't counted in the any of
hit_kb_out counters so I would still see the discrepancies in figures
between web-polygraph and a single instance squid (never had a need to
check before now).

2). Excluding the fact that headers may not be counted, does the
formula I'm using sound like the correct way to calculate hit % with a
multi-instance setup

3). From the 3.2 wiki page - http://wiki.squid-cache.org/Features/SmpScale
        Currently, Squid workers do not share and do not synchronize other
resources or services, including:
        • object caches (memory and disk) -- there is an active project to
allow such sharing;

Can 3.2 workers be configured with other workers as siblings to make
use of their cache.
Received on Thu Dec 16 2010 - 04:37:21 MST