Re: Async I/O on IRIX 6.x?

Alex Rousskov wrote:

> More precisely, has anyone gotten asynchronous I/O to work better
> than synchronous I/O on IRIX 6.x? :)
>
> If there are async/IRIX gurus out there, here is what we have:
>
> Workload:
> 40 simulated clients, each submitting requests as soon as
> they get reply,
> 1 simulated server
> no hits
> all requests cachable
> no persistent connections
>
> The configuration gets about 500req/sec thru a null-proxy. (Null-proxy
> runs on the same machine and just forwards requests to the server and
> replies to the clients).
>
> When Squid 1.2 replaces our null-proxy and starts with an empty cache,
> we get about 30req/sec for sync-io and 12 req/sec for async-io.
>
> All these numbers are preliminary, but the sync/async difference is
> already obvious and disturbing.
>
> Async-io always complains about NUMTHREADS being too low (currently
> 128, tried 16, 32, and 64). The distribution of waiting threads may
> look like this:
>
> Queue Length = 2278
>
> # %
> OPEN 378 16.59
> READ 0 0.00
> WRITE 804 35.29
> CLOSE 1096 48.11
>
> #reads is always 0. #close is always close to 50%.

Ugh. This does not look right for 40 clients. Far to many outstanding
requests. 378+804=1182 active files...

The distribution mostly what's expected, but the numbers are large.

> Any clues? What tools are async-people using to debug threads
> performance?

I mostly use tcp-banger when comparing async/sync performance. I do
however lack a fast server to fill the cache, and the most tests
I have done is cache hits.

I haven't seriously tested async-io any async-io in the last mont
or so.

/Henrik
Received on Tue Jul 29 2003 - 13:15:53 MDT