Date: Thu, 11 Apr 1996 19:35:25 -0500 (EST) From: "John S. Dyson" <toor@dyson.iquest.net> To: luigi@labinfo.iet.unipi.it (Luigi Rizzo) Cc: se@zpr.uni-koeln.de, bde@zeta.org.au, dutchman@spase.nl, freebsd-hackers@FreeBSD.org Subject: Re: HDD cpu usage (IDE vs. SCSI). Message-ID: <199604120035.TAA05654@dyson.iquest.net> In-Reply-To: <199604111834.UAA15330@labinfo.iet.unipi.it> from "Luigi Rizzo" at Apr 11, 96 08:34:18 pm
next in thread | previous in thread | raw e-mail | index | archive | help
>
> Sorry for not trimming the message too much but the numbers sound
> strange.
>
Yep... If you assume that the purpose of the benchmark is to measure
total drive perf. Actually, it is meant to measure the interface overhead.
>
> but the numbers for the NCR probably suggest that data ought to be
> averaged over a larger number of experiments.
>
True, but it still appears that the NCR/HAWK overhead is high. I would
tend to suspect the HAWK more than the NCR -- but I simply do not know.
>
> Second, I guess you are measuring the SCSI/IDE/EIDE bus transfer speed,
> rather than the disk transfer speed, isn't it ?
>
Yes. The purpose of the benchmark appears to be to measure the interface
performance -- not necessarily the drive itself.
>
> Also, how comes that a 2X CDROM has such a low speed ?
>
Apparently the SCSI overhead of the TOSIBA XM-3401TA is very very high.
>
> Third, what do you mean by "Standard IDE" vs EIDE ? Isn't it rather an
> ISA vs. VLB/PCI comparison ? I have tried both a WDC540 and a WDC1.6GB,
> and the 540 gets a maximum speed (with iozone or bonnie) of some
> 2.2MB/s at most, no matter how fast is the interface or the system.
>
It is standard IDE timing vs. MODE 3/4 timing. PCI interface isn't important
here -- IDE is kind-of an ISA bus with perhaps modified timing. Below
is the performance on my WDC540H (the good one with the big cache), using
bonnie and MODE3 +- timing.
-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
100 1663 35.2 2797 11.9 1017 5.2 2074 33.3 2868 7.5 62.5 2.7
> Instead, the WDC 1.6GB (as you also experienced) is rated at 5.5MB/s on
> the same system).
>
> > Standard IDE W/WDC540H
> > Command overhead is 436 usec (time_4096 = 2001, time_8192 = 3565)
> > transfer speed is 2.61859e+06 bytes/sec
> > Command overhead is 436 usec (time_4096 = 2002, time_8192 = 3568)
> > transfer speed is 2.61586e+06 bytes/sec
> >
This says NO MATTER what the drive can do, the max xfer rate is 2.6MBytes
per second using regular IDE timing. At that rate, it is likely that there
will be 100% CPU utlization just for the I/O xfer :-(.
> > EIDE W/WDC540H
> > Command overhead is 375 usec (time_4096 = 814, time_8192 = 1254)
> > transfer speed is 9.31994e+06 bytes/sec
> > Command overhead is 368 usec (time_4096 = 800, time_8192 = 1231)
> > transfer speed is 9.49095e+06 bytes/sec
> >
This says that the WDC540H is capable of much better than IDE timing. Note
that it is likely that the CPU overhead is much much less using enhanced timing,
even though the drive is not capable of much better than 2.7-2.8MBytes/sec.
There are lots more CPU wait states when transferring in standard IDE mode
than in MODE3.
> > EIDE W/WDC 1.6GB
> > Command overhead is 217 usec (time_4096 = 513, time_8192 = 809)
> > transfer speed is 1.38444e+07 bytes/sec
> > Command overhead is 196 usec (time_4096 = 502, time_8192 = 808)
> > transfer speed is 1.33987e+07 bytes/sec
>
This says that the EIDE 1.6Gbyte drive has very quick command overhead, and
can transfer between the drives buffer and CPU at 13MBytes/sec (perhaps
more, but that is what my machine does with it.) The actual platter
transfer rate through FFS is shown below:
-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
100 1816 39.5 5442 25.8 1881 9.2 2271 37.6 5628 13.4 89.6 3.6
BTW, next time I want to driver-hack, I plan to work on Garrett's submission
of the DMA code for Triton EIDE. It would be interesting to see (on an
experimental basis) if the CPU loading is significantly decreased further.
John
dyson@freebsd.org
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199604120035.TAA05654>