Re: 32 bit I/O DMA and IDE drives

I have to humbly disagree with Dirk, on this subject.

-u can be a dangerous option. And in extensive benchmarks, I've been
unable to find a measurable difference between using it and not using it
under Squid loads.

DMA is not always enabled by default on new hardware. Some UDMA66
controllers do require it to be turned on during system startup. And
some of the of newest UDMA 66 drives will not work in DMA mode at all on
some of the newest UDMA 66 controllers under Linux (Via 586/686 UDMA 66
controller chips and Seagate and Quantum 7200 RPM drives exhibit this
phenomenon).

-m16 also has only negligible impact on performance. 1-2% in fact.
Possibly statistically meaningless. It may be a big win for some types
of loads, but for Squid it's pointless to worry over.

DMA does make a big difference. In my tests, throughput can drop from
60 reqs/sec to about 40 reqs/sec, when a drive is not using DMA mode.
Of course, this is with a very CPU hungry async i/o compile of Squid, so
with DiskD or plain UFS, the difference is somewhat smaller--The current
speed limit of DiskD while maintaining good hit ratio, 45, drops to
below 30 when DMA is disabled.

32 bit I/O makes little difference in my tests.

My $.02

Dirk Moerenhout wrote:
>
> Normally the most important hdparm parameters are -u and -m when it comes
> to speed. You really need -u 1 to work if you want to have good
> performance with an IDE disk. For -m use something like -m 16.
>
> If DMA is not used by default something fishy is going on and you must be
> using outdated parts. You could still try it but any recent kernel will
> turn on DMA if hardware allows it.
>
> Enabling 32bit I/O gives as good as nothing as a gain.
>
> Note that -u 1 may cause problems but at the same time it is the most
> important tuning you need. And of course try to get some SCSI :-)
>
> Dirk Moerenhout ///// System Administrator ///// Planet Internet NV
>
> On Sat, 5 Aug 2000, John Deschamps wrote:
>
> >
> > For now I'm stuck with using an 8GB IDE drive for RH 6.1/Squid
> > 2.3.STABLE2. Will enabling 32 bit I/O and DMA via hdparm improve squid
> > performance and/or will it introduce (in)stability issues?

                                  --
                     Joe Cooper <joe@swelltech.com>
                 Affordable Web Caching Proxy Appliances
                        http://www.swelltech.com
Received on Mon Aug 07 2000 - 13:01:22 MDT