Re: [squid-users] Squid response time

tis 2003-03-04 klockan 10.21 skrev Wei Keong:
> > On Tuesday 04 March 2003 02.47, Wei Keong wrote:
> > > Squid 1 has no cache, only converting proxy request to web
> > > request...
> >
> > What does cachemgr say about hit rate on Squid 1?
>
> Squid 1 has 128MB of cache mem... Req Hit about 15%, Byte Hit about 5%

Which is quite significant good when it comes to service times.. memory
hits are very quick, and 15% of the request to squid 1 is handled
"instantaneously".

> I obtain the MST from cachemgr, and plot it onto RRD graphs. The HTTP
> Request (All) and Cache Misses for Squid 1 is clearly lower than 2.

Which again makes sense as a percentage of the cache misses on squid 1
will be cache hits on squid 2, which should be quicker than a cache
miss..

The MST for cache misses on Squid 1 should approximate the MST of all
requests on Squid 2 as these two values represent the same requests
(assuming there is no other users of Squid 2).

> > > Supposing we are downloading a file, does Squid writes into access
> > > log when it receives the response header or when it receives the
> > > whole file?
> >
> > Squid access times are calculated from the time Squid has received the
> > request headers to where it has written the full response to the
> > client socket.
> >
> > Another thing it can be is TCP windows. Most "server" OS:es have much
> > larger TCP windows than microsoft desktop OS:es...
>
> This shows that MST should be a valid measurement of response time...

The Squid MST is always an approximation. It does not include

  a) The time it takes for the client to open the connection and send
the request

  b) The time it takes for the reply to be sent to the client after it
has been sent to the TCP socket by Squid.

Both of these is very hard for Squid to measure, and needs to be
measured on the client to get reliable values. If the clients are on
slow dialup connections the time in 'a' and 'b' can be significant, but
for LAN clients both approximate 0.

> Btw,
> wget test verified that Squid 1 takes longer to download...

How big difference, and what was logged in access.log on the two
servers?

System wide cache misses will quite naturally be marginally slower in
setup time when going via two proxies than one as there is one more
application hop the request has to go via, but batch throughput should
be the same once the download has get started and on average/median you
should gain quite a bit in response time from the cache hits..

Regards
Henrik

-- 
Henrik Nordstrom <hno@squid-cache.org>
MARA Systems AB, Sweden