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

Alex Rousskov wrote:

> I have spend a lot of time fixing the async-io stuff on IRIX.

What? NLANR looking into async-io? Have you changed your mind on this
issue? You was not very positive about threading and async-io when we
spoke in Manchester.

> On the other hand, what happens on a loaded proxy with a good network
> connectivity? Imagine fast pumping network working side-by-side with a
> relatively slow disk... Pretty soon the Squid disk queues exceed any
> reasonable limits, while the number of pending open(O_CREATE) requests
> continues to grow.

Oops.. I thought that store waited for the open to complete, but I see
now that it doesn't. The only delay there is perhaps when the number of
filedescriptors runs low and Squid stops accepting new connections. This
could cause a large number of problems as it does not only overload the
disks, it also causes the memory to fill quickly causing many swapouts
to be aborted when their in-memory object is ejected due to lack of
cache_mem..

> Thus, after fixing the IO code, we got what we wanted -- a very fast Squid.
> Unfortunately, that "speed" has to be dealt with. Currently, we monitor the
> queues and do not cache objects while the queues are "too long". The
> performance of this disk admission control algorithm (as of the async-io code
> itself) still leaves a lot of space for improvement, IMO. Working on it.

Maybe we should spend some time on having these queues limited per disk
(or cache_dir). Many locations may have disks of different speeds,
especially after upgrading their cache (newer disks faster and larger
than the older ones).

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