[…]

> client<->squid :  31Mbps.

[…]

> Squid<->server : 28Mbps
> 
> Total: 59Mbps.
> 
> Which is slightly higher than the known "good" performance limit for 
> Squid-3.1. Which is up to ~50Mbps, tuning both in Squid and the system can 
> reach around 100Mbps IIRC. But that sort of numbers you are looking at 
> specific traffic types as well.

Ah ! Thank you very much for this valuable figures.
50Mb/s is not what I was really hoping/expecting. I was naively thinking that 
Squid can handle multi 100Mb/s out of the box (at least with no ACLs, and other 
fancy stuff)
We have 2x100Mb/s Internet connections in front of this proxy.

>> select_loops = 2316.561716/sec
> <snip>
>> syscalls.sock.reads = 1316.883520/sec
>> syscalls.sock.writes = 1770.096831/sec
> 
> It looks like sending and receiving data in ~2KB chunks, just over one packet 
> per cycle.
> 
>> syscalls.sock.recvfroms = 58.203913/sec
>> syscalls.sock.sendtos = 30.974986/sec
>> cpu_time = 117.934176 seconds
>> wall_time = 300.048557 seconds
>> cpu_usage = 39.305030%
> 
> CPU usage at 40%.
> 
> Squid-3.1 does use around 30% CPU to push data at line rate, which is what 
> this set of counters seems to be showing. We have done a bunch of work aimed 
> at improving things all over Squid in the more recent releases, if you want 
> to assist the developers tracking this down please upgrade to the latest 
> release before delving too deeply into the analysis as all fixes need to go 
> in through 3.HEAD anyway and you maybe repeating already done work (only 
> might there is still CPU issues hiding in the latest development code).

Point taken, I'll do that.

> What do things look like when its at or close to 100% CPU?

Basically, some users complain about browsing slowing down. But the main 
complaint is about download speeds of "large" files (~50MB -> GB) (but this can 
be explaind by the natural limits of Squid)

> Also any sign of the OS swapping virtual memory when the CPU climbs? that is 
> a known performance killer for a number of reasons we can't fix easily.

Already checked this area, no swapping involved.

> Amos

Youssef

Reply via email to