> I have no idea what the point of parallel hdparm's is.
Recently I wanted to know the performance of a system assuming
infinitely fast disks but using the actual IDE controllers, PCI
subsystem, etc. hdparm is indeed the wrong thing for this: it
does either actual device reads (at the speed of the disk)
or buffer-cache reads (at the speed of the memory system, not
including the IDE bus, DMA controller, and PCI bus).
A close approximation to the right thing, I think, is to use
a raw device. If you read only from the first, say, 512 kByte
of the disk, the on-disk cache will supply the data. (I think
I used lmdd, since you could tell it to repeatedly access the
same region.)
For ATA-66 on the Promise PCI card, I think it topped out at around
50 MBytes/sec for a single interface; with both interfaces aggregate
speed was around 95 MBytes/sec. (440BX motherboard.)
So for a future disconnect-reconnect-capable IDE driver, we ought
to see these sorts of numbers for IDE buses with two devices in
actual, realistic situations.
I also tried i820 with the supplied ATA-66 I/O hub; similar
numbers if I remember right. The attractive part might have been
that the disk paths wouldn't contend with PCI devices, but after
using some PCI NICs using significant bus bandwidth in this scenario,
the picture isn't so rosy: the "hub" business still shows bottlenecks
to memory.
Cheers,
Peter Monta pmonta@terayon.com
Terayon Communication Systems
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Fri May 19 2000 - 21:48:31 EDT