[Prev][Next][Index]

Re: Hardware for Internet providers




El-Hanany Yuval writes:

 > This thread seems to advocate the use of PC's instead of real work stations.
 > I must say that my experience, although quite pleasant with linux, seems to
 > disagree with this approach. At my work place, we had a PC pentium 32MB
 > running Solaris. On the paper, it should have been our strongest machine at
 > the time. However, it was OK as long as you stayed within the memory limits,
 > once you exceeded and started swapping, it was a lost cause. It served as an
 > X Terminal until we got our orders of Sun Workstations.

Solaris is a dog - huge binaries, huge memory requirements.  The same
machine running linux would rarely swap much.  Also, you don't say
what kind of disk - IDE or SCSI, and what kind of bus - ISA, EISA,
VESA, or PCI?  If you're talking about IDE+ISA, then yes, you probably
won't have great I/O throughput and hence will not have great swapping
performance.

 > At home I've got a 486 16MB SCSI linux, and while it's great, swapping is
 > far from enjoyable and is not comparable to any of the WorkStations brands
 > I've worked with (RS/6000, Digital, Sun, SiliconGraphics).

Firstly, what kind of 486, how old, what bus?  If it's a 486/33 with
ISA bus, then I'm not so surprised.  The machine, and the disks in
particular could be quite abit slower than mid range workstations,
possibly even slower than low end workstations that are currently on
the market (as opposed to low end workstations that are extinct, such
as sun 3s and sun 4s, which would probably be much slower than this PC).

Secondly, what Linux kernel?  Does it have the kswap patch installed?
The swapping algoritim in the 1.2.x kernels isn't especially good.  I
hear that the kswap patch makes swapping *alot* faster.

Thirdly, what kind of swapping are you talking about?  Lots and lots
of little processes competing for run time, or one huge 20mb number
crunching program which is constantly bouncing through its data space?
The latter's performance would go to hell on *any* machine with 16mb
of RAM.

Fourthly, a fast hard disk on a fast SCSI card on a PCI bus on a good
Pentium motherboard under Linux will give *very* good disk I/O speeds.

Here're some iozone results for my P120 with an *IDE* hard disk on a
PCI bus:

Run 1:

   Writing the 40 Megabyte file, 'iozone.tmp'...28.520000 seconds
   Reading the file...19.960000 seconds

   IOZONE performance measurements:
	   1470653 bytes/second for writing the file
           2101354 bytes/second for reading the file

Run 2:

   Writing the 40 Megabyte file, 'iozone.tmp'...28.780000 seconds
   Reading the file...19.770000 seconds

   IOZONE performance measurements:
	   1457367 bytes/second for writing the file
	   2121549 bytes/second for reading the file

Here're the results of running iozone on a Data General AViiON with 1
gb of RAM (yes, RAM), and 8 mc88110 processors.  Note that these DGs
are known for their (excellent) I/O performance.

Run 1:

   Writing the 40 Megabyte file, 'iozone.tmp'...22.570000 seconds
   Reading the file...40.160000 seconds

   IOZONE performance measurements:
	   1858353 bytes/second for writing the file
	   1044398 bytes/second for reading the file

Run 2:

   Writing the 40 Megabyte file, 'iozone.tmp'...20.540000 seconds
   Reading the file...16.390000 seconds

   IOZONE performance measurements:
	   2042017 bytes/second for writing the file
	   2559062 bytes/second for reading the file

Note that my P120 with IDE drives on a PCI bus is not lagging too far
behind this huge behemoth of a unix beast!

Add to this the kswap patch, and a fast SCSI disk and you'll have a
machine with quite fast swapping.

Furthermore, I've heard reports of people pushing 4 megabytes/second
through the filesystem on Linux machines with ncr SCSI built into the
motherboard.  This is already getting close to the theoretical
throughput limits on most current hard disks.

 > As far as I know of the internals of your average PC that shouldn't be
 > surprising since it still suffers from the drawbacks of the original PC
 > (although that may have changed already in the Pentium machines).

Not really.  200 mhz DEC alphas are also running PCI bus.  But even
the ISA bus is underappreciated.  It can run at about 8mhz with a 16
bit data path, yielding a bandwidth of about 16 megabytes per second.
I don't know of any disk drives that can come close to saturating
this.  Today's fast disks can still only put out a max of something
like 4-8 megabytes per second.  Disks rarely saturate an ISA bus.
What really kill it is video - moving windows around, displaying
MPEGs, etc.

The problems with the ISA bus are that

   1 - it can take alot of dedicated processor time, and
   2 - everything else (I/O interrupts, video, etc) isgoing through it
       and grabbing bandwidth, bus timeslices, and dedicated processor
       time.

This is what kills it for making a really all around fast machine.

Of course, like I said, using SCSI & PCI bus pretty much alleviates
the problem, so any good high end PC (PCI, SCSI, Linux, kswap patch)
should perform quite favorably compared to the workstation offerings
of IBM, Digital, Silicon Graphics, Sun, ...  Of course, for number
crunching, a 133mhz pentium is still about 4x slower on floating point
than a 200mhz alpha...

-- 
Dr. Harvey J. Stein
Berger Financial Research
abel@netvision.net.il (temporary address)