Re: Slow disk performance (was: Re: install help 6400/200)
Michael, thank you for the corrections. I have some comments may also be
useful for those considering an old PowerMac as a Debian platform.
Now I am building a Debian high performance server using an 8500 as the
base system, and I hope the experience help people trying to use Debian
on the PowerMac for similar tasks.
You are absolutely right, only 32 bit PCI. I am sorry for the imprecise
comment. I got confused because for most high performance peripherals is
recommended to use slots 1 and 4. These are the first on each of the
9500 PCI buses, and ensure better throughput (I do not know the reason.)
This same machine supports a ten-fold increase in transfer rates using
newer SCSI appliances. The 9500 provides 6 PCI slots and two of them are
64 bit wide.
Errmm.. no, the 9x00 series have no 64-bit slots. Those were the
first-generation PCI machines. You're a generation too far :-).
Just add a good SCSI card (Ultra-2 LVD or Ultra160) and a
newer disk to see real (not bus nominal) transfer speeds of 25-40 Mb/s.
Or an IDE card, for that matter. I get close to 15 MB/s on a (now rather
old) 10 Gig Maxtor 7200 rpm disk with a Promise Ultra/66.
You are right. I did not mention IDE because it usually blocks the
system for too much time during disk operations.
I have no experience with IDE on PowerMacs, but on PCs a single
large-file copy is a sure way to slow down the system to a crawl. This
is not a problem on most workstations, but for a server with more then
minimal traffic it may be a real concern.
Here an advice based on system usage: for most workstations an IDE drive
will deliver cheap performance. For a budget-minded server it may be
better to use a cheap scsi drive, even if its slower than the IDE. The
scsi system will perform better when several tasks need to cohexist and
do lots of disk operations.
Of course those who can afford Ultra2 or Ultra160 mirrored-controller
striped-disk systems will experience stunning results for any kind of
Some disk controllers even provide striping features, letting you read
file fragments in parallel from different physical disks. Some setups
deliver sustained 130 Mb/s read data rates.
You will never get that much (supposing you mean 130 MB/s). For one, a
single disk today will not go above maybe 40 MB/s. Anything above can
only be obtained with expensive high-end RAID systems; and then you're
starting to hit the limits of your storage attachment system, be it
SCSI/320, Ultra/133 for IDE, or Fibre Channel (1 or 2 Gbit/s).
Yes, with a high end single disk/controller, 30/40 MB/s is the norm.
I sounds like madness, but an 9500 can handle mirrored controllers on
its dual PCI buses. Some performance addicted have tried 4-disk setups
where each controller handle a striped volume set with stunning results!
On limiting the madness: 32bit PCI can not handle Ultra160 throughput
with striped volumes. A pair of high performance disks connected to a
single controller will certainly saturate the PCI bus on read operations
(peaks of 60-70 MB/s).
I am trying to reach these performance levels on my 8500, but I am still
doing reserch on Linux (Debian) compatible hardware.
Right, 80MB/s on each bus. The 9500 has 2 "bandit" chips, each handling
one of the PCI buses.
Coming back to the 9500, you'll also hit the limit of that system's PCI
buses, which can do (according to Apple's doc: see technote 1008) a
maximum of 80 MB/s under optimal conditions.
All but the 9500 (with the soldered cache) and the mach5 9600s (where
the L2 cache was dropped because the chip comes with a large on-chip
cache, like the G3).
The 9500 has an additional problem, because its 512 Kb L2 cache is soldered
in the mainboard and can't be replaced, but usually it is a good quality
Sure about that? All the other first-gen systems have their L2 cache on
Michel Lanners | " Read Philosophy. Study Art.
23, Rue Paul Henkes | Ask Questions. Make Mistakes.
L-1710 Luxembourg |
email email@example.com |
http://www.cpu.lu/~mlan | Learn Always. "
To UNSUBSCRIBE, email to firstname.lastname@example.org
with a subject of "unsubscribe". Trouble? Contact email@example.com