Re: [squid-users] relation between http req/sec and no of simulatneous accesses

On Fri, Oct 12, 2001 at 01:45:52AM -0600, Alex Rousskov wrote:
> On Thu, 11 Oct 2001, khiz code wrote:
> > my cache could go upto abt 70 req/sec
> > what i would like to understand is that 70 req/sec corresponds to abt
> > how many simultaneous client connections ..i.e how many clients can the
> > cache serve at any point of time
> >
> > Robin stevens had given a figure of 2500 simultaneuos access for
> > 150 req/sec so cud i apply simple math logic to arrive at a
> > corresponding statistic for 70 req/sec which turns out to be
> > roughly 1000 simultaneous accessess ???
>
> You cannot apply "simple math logic" here unless your response times
> are the same as what Robin Stevens observed.
 
Moreover response times here vary quite a bit, especially on cache misses.
Median miss times are 250-300ms at quiet times rising to over 500ms during
busy periods, probably due to congestion at the parent caches and on links
into external networks. Maximum connection times can be several hours
(hosts on dialup/ISDN performing large downloads, people downloading
1GB+ files of research data); cache hits to well-connected machines can be
served in a few milliseconds.

The number of clients you can cope with at a rate of n requests/sec is
going to depend a lot on usage patterns and how much each client is going
to use the web - a large network of machines making light usage and a small
network of machines making heavy use will require similar resources. The
speed of your link is also an issue.

For what it's worth, we see around 14000 IP addresses accessing the proxies
per day (the true number of active clients will be higher as some networks
employ NAT gateways, DHCP, etc), but with very significant usage well
beyond the standard 9-5 working day (usage at midnight is still over 50% of
the afternoon peak). A much smaller number would be active at any one time
(judging by netstat output it's around 1000 hosts across the cluster at a
given instant) or indeed within a 1 hour period. Patterns on a corporate
network would no doubt be very different.

-- 
--------------- Robin Stevens  <robin.stevens@oucs.ox.ac.uk> -----------------
Oxford University Computing Services ----------- Web: http://www.cynic.org.uk/
------- (+44)(0)1865: 273212 (work) 273275 (fax)  Mobile: 07776 235326 -------