[linux-elitists] Performance issues with swap file vs. swap partition

Jason Spence jspence@lightconsulting.com
Thu Nov 6 11:54:27 PST 2003


On Thu, Nov 06, 2003 at 01:49:25PM +0100, Eugen Leitl wrote: 
> On Thu, Nov 06, 2003 at 01:28:14PM +0100, Ragnar Hojland Espinosa wrote:
> 
> > AFAIK, the performance difference on modern kernels is negligible.
> > That said, if you have a swap partition on a separate IDE channel or
> > SCSI disk, you won't starve the I/O bandwidth by swapping (which would
> > probabily happen if you have a swap file..)
> 
> Speaking of more spindles, does SATA address the usual IDE problem (system
> has response hiatus (keyboard, audio) during intense IDE activity)? Sure,
> SCSI, I know, don't have the budget for that one at home. No way how I could
> pay half a TByte worth of SCSI drives.

Huh?  I never noticed there was a problem with IDE in general.  Many
times someone will forget to turn DMA on (or cannot due to older
hardware) and then I'll see performance problems, but I've never had
any issues even with multiple disks going once DMA transfers are
enabled.

I have a SATA system here with a 10K cheetah from Western Digital and
dd if=/dev/hde of=/dev/null gives me 57 MB/sec worth of data and ~30%
of system usage for the task on a HT enabled P4 @ 2.4GHz.  No
noticible latency issues whatsoever, but I'm running redhat 9 which
includes something like 300 patches in their kernel, including low
latency and preemptive patches.

I can't post a comparable SCSI benchmark because all my machines with
SCSI go through a RAID card and so I get ~1% system usage for the dd
task.  They're also running FreeBSD...

I do recall streaming anything to/from (SCSI) disks on my SparcStation
20s at media rate (to a file on the filesystem, not the block device)
would peg the CPUs, and was a bottleneck in the case of an IPX I had
lying around somewhere.

For whatever reason, a real low end IRIX machine (an Indy) had much
lower CPU utilization compared to the SparcStation 20s even with the
exact same disk array attached.  It was still a pretty high absolute
number, but lower compared to the SS20s.  xfs vs ufs?  MIPS vs SPARC?
Dunno.

> Is responsiveness visibly increased in general in the 2.6 kernel
> family (haven't run one yet)?

I can't run them "on the metal" - ACPI support is still really damn
broken for my bitch boxes.  In fact, the way they're planning on
supporting ACPI makes me wary of running 2.6.x at all on pretty much
any of my machines, since a bad DSDT or a flaw in the interpreter can
toast your hardware, and I tend to be running bleeding edge stuff
people lend to me for my 3d work; those machines are more apt to have
ACPI tables that would get misinterpreted by the AML interpreter, I've
found.

I'm seriously considering switching entirely to FreeBSD and Windows
for the duration of the next release cycle, since they both have the
support I need to do really weird stuff like hot swap PCI cards and
make servers suspend-to-disk for site relocations.  But on the other
hand, there's no SMART or PFA support in FreeBSD, NVIDIA driver
support is several months behind, and Windows is expensive.

In vmware, there wasn't anything immediately noticible when I tried
2.6, but I didn't have CPU contention issues because those machines
were all really fast or parallel and I wasn't doing anything that
would really beat up the scheduler.

-- 
 - Jason       Last known location:  11.0 miles northeast of MOUNTAIN VIEW, CA

Re graphics: A picture is worth 10K words -- but only those to describe
the picture.  Hardly any sets of 10K words can be adequately described
with pictures.



More information about the linux-elitists mailing list