yes but for us, latency is also an important factor. increasing max object
size means more cache storage space is being consumed by bigger objects,
which might get hits less too often. we have object hit ratio at around
50% and our byte hit with current setting is around 33%. with larger
objects in cache, byte hit could probably be raised to above 40%, can't
it? but the request hit drops too i suppose. its a tradeoff.
about 50% byte hit. latency is OK
On Mon, 27 Nov 2006, Wojciech Puchar wrote:
I think 5.7mbits/sec traffic with 18req/sec per cache (about 54req/sec for
3 cache) makes around 108kbit/request. still is quite big though. we allow
objects upto 8MBytes to be cached.
strange that you eliminate hits that saves the most traffic by limiting
object sizes.
i use 10GB max size.
On Mon, 27 Nov 2006, Adrian Chadd wrote:
On Mon, Nov 27, 2006, Manoj Rajkarnikar wrote:
request per second per cache is around 18 at peak time. traffic flow
combined to all three cache is around 5.7Mbits at peak. do you think this
cpu usage is normal for the provided load ??
Not for 18 per second; but 5.7 mbit/peak at 18req/sec is averaging ~ 300k per
reply. Thats quite big. :)
Did you compile with epoll() for linux or kqueue() for FreeBSD? Or, hm, can
you just paste ./squid -v here?
Thanks,
Adrian
--
Manoj Rajkarnikar
System Administrator
Vianet Communications Pvt Ltd
Pulchowk, Lalitpur, Nepal.
(PH)977-1-5546410
--
Manoj Rajkarnikar
System Administrator
Vianet Communications Pvt Ltd
Pulchowk, Lalitpur, Nepal.
(PH)977-1-5546410
