Alfonso Correas <acserra@ibm.net> writes:
> Hi squidders,
>
> We need some kind of advice, because a customer has requested us a big
> behemoth of Squid configuration. They requested a configuration capable
> of handling +800Hits/second on peak times.
>
> This is the hw, that we think right.
>
> Pentium II 400Mhz.
> Intel 440BX motherboard.
> 512Mb SDRAM.
> DPT PM2144UW in RAID-5.
Very bad move. This key thing to remember about RAID-5 is that the
write performance sucks. The other key thing to remember about squid
is that it writes almost everything that it fetches to disk...
Our fairly happy squid (sustained 50 requests/sec, peaks of 150 (5
mins averages)) writes around 1/2 a meg / sec to disk sustained.
> Quantum Atlas II 9.1Gb UW (5+1) in RAID cabinet.
> DEC 21140 PCI 100Base-T LAN card (2).
>
> OS: Linux/FreeBSD¿?
>
> We are experimenting with a dual LAN card to handle the incoming
> requests via one interface and forward the requests via other interface,
> to alleviate the interface congestion.
If you're running full duplex, that won't be a consideration I should
think. And keeping it on the one card will reduce the number of
interrupts.
> Can be possible to disable logging to alleviate the overhead rather than
> writing in other disk? What about performance tricks ala' NNTP servers?
Yes, disabling logging is possible, but it probably wouldn't save you
much. More ram would be a better bet.
Note that it's possible to run a smaller cache MUCH faster than a
larger cache (the disk cache will cache more effectively).
Michael.
Received on Sat Apr 25 1998 - 21:21:54 MDT
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:39:57 MST