Re: 1.2.20beta async-io

Dirk Vleugels wrote:

> How much of a performance improve in comparison to a 1.1 squid? Does the
> threading pay of?

This is best answered with some hard numbers from my test machine:

                                random
                fill(1) read read
Squid 1.1.21 : 35 42 34
Squid 1.2b20-1: 42 105(2) 60

(1) Cache fill probably injured a major penalty from the local HTTP
server sharing one of the disks.
(2) Sequential read using 1.2b20-1 was CPU bound.

The test machine is a

Pentium 133, 32MB memory, RetHat Linux 5.0
2 IDE drives (QUANTUM FIREBALL_TM3200A, FUJITSU M1636TAU)
local HTTP server (Apache)
No real network. All running local on one machine.

The test data consists of ~5500 objects with a total size of ~30MB.

Both Squid versions used identical squid.conf settings (apart from the
syntax difference).

Squid 1.1.21 is a stock Squid without any (or very small) modifications.
Squid 1.2b20-1 is my development version of Squid 1.2b20-1. See
http://hem.passagen.se/hno/squid/ for a complete listing of all changes.
NUMTHREADS was set to 64.

All three tests is measured by tcp-banger. The fill test is utilizing 10
simultaneous connections and the read tests uses 60 (-n option to
tcp-banger).

sequential read == in the same order as used when filling the cache.
random read == in random order.

---
Henrik Nordström
Sparetime Squid Hacker