On 13/01/2013 5:44 p.m., John Joseph wrote:
Hi Loic
Your feedback was quite useful, I was able to come up with some values after checking your configuration.
If by 3 years my users base increase to 4000 no, then will it be a over kill if I go with 32 GB ram.
4000 "users" (I assume that means, 1 client software == 1 user) all
simultaneously downloading (+1 client I/O buffer) large objects (>256KB)
which are *all* TCP_MISS (+1 server I/O buffer) and being REQMOD *and*
RESPMOD ICAP filtered (+2 network I/O buffers) will consume ~5GB of RAM
in Squid.
+ 1x client I/O buffer, 64KB
+1x server I/O buffer, 64KB
+2x ICAP I/O bufers, 128KB
The likelyhood of that happening is relatively low if you are an ISP.
Far more likely is that you will have a mix of HIT/MISS, and notbe ICAP
fitering some or most requests. Squid can operate happyly with a few
hundred MB of RAM
The big RAM consumption comes from cache_mem, which is Squids in-memory
filesystem for cache storage of high demand objects. That can consume as
much or as little RAM as you can throw at it, up to a limit which is
higher than most can afford to purchase yet.
As has been suggested earlier, *drop* the idea of "users" when
calculating HTTP requirements. Users are irrelevant. One single user can
completely max out a Gbps ethernet connection, and several thousand
users can happily co-share a 56Kbps uplink. The traffic request rate and
size are the key details for capacity planning, followed by the amount
of processing components you are going to be performing on that traffic.
IMO, unless you are expecting to face some particularly unusual
situation like hundreds of thousands of users or very high traffic rates
the commonly available hardware can handle the traffic easily. Look for
good I/O speeds with low latency on disk hardware, high write speeds for
any SSD hardware planned, and the rest can be governed by your available
budget.
----- Original Message -----
From: Loïc Blot
Hello Joseph,
I use a Dell R320 (2, because failover), under OpenBSD 5.2 with 16GB RAM
and Two Intel PRO/1000 PT (82571EB) (Broadcom 5720 isn't supported).
I have 500-600 users/smartphones and 1GB of WAN bandwidth.
To improve perfs (and this explain why so many RAM), i
move /var/squid/cache to mfs (memory file system), and i use 4G "disk"
cache and 3.5G memory cache. Perfs are very great.
I'll bet. You realize they could possibly be even better by using a
64-bit system and eliminating the "disk" cache? Squid swaps objects from
memory cache to "disk" cache and back again during regular operation. If
you have enough RAM to eliminate that, why bother forming a
configuration that keeps the overheads?
Amos
