Re: [Fwd: Re: [SLUG] IOWait definition]

2008-10-09 Thread Adrian Chadd
iostat can be a bit special.I -think- under linux its the amount
of time spent waiting for pending disk IO to complete. Now, 
some chipsets and their drivers seem to spend a lot of time in IOWAIT
compared to others. The traditional difference was polled vs dma'ed
disk IO - with polled IO, the driver would submit a request and then
sit there and wait for it to complete. That time spent waiting was
IOWAIT. DMA'ing controllers would spend less time in IOWAIT because
they'd submit a request, kick it off, it'd happen in the background,
and then the kernel would be notified when it completed. Almost no time
was spent in IOWAIT - just the time scheduling the DMA commands and
handling the response.

Now, I've seen some low-ened SATA chipsets on some reasonably speccy
hardware (eg some of the nvidia-driver Sun workstations) with large
amounts of IOWAIT time. I don't know if its a driver thing or a hardware
thing (or both), but there are definitely issues.

Now, you could go groveling through the kernel to try and figure out
whether IOWAIT includes nfs activity (which I believe would be
the disk IO related to NFS, but it could be other things I guess) and
first see if doing userspace disk IO does a lot of IOWAIT. Grab bonnie
or some other hard disk throughput testing thing, run it, and see how
much time is spent in IOWAIT. If you've got almost no IOWAIT time when
doing it locally but IOWAIT time when doing it over NFS, you could be
right. If you have IOWAIT time on both, I'd poke the disk/controller/
driver combo. 

2d,



Adrian


On Thu, Oct 09, 2008, David Kempe wrote:
 Grant Street wrote:
 
 
 I have a machine with a good proportion of IOWait 20-30%. It does have
 local disks and it performs operations on NFS mounts. I just wanted to
 be sure if IOWait includes NFS activity or not. I also want a way if it
 is NFS to be able to say for sure if it is a bottleneck on the nfs 
 client or server. NFS is not a linux machine so visibility is not 
 allways the best.
 
 dstat might help you correlate stuff.
 http://dag.wieers.com/home-made/dstat/
 
 dave
 -- 
 SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
 Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html

-- 
- Xenion - http://www.xenion.com.au/ - VPS Hosting - Commercial Squid Support -
- $25/pm entry-level VPSes w/ capped bandwidth charges available in WA -
-- 
SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html


Re: [Fwd: Re: [SLUG] IOWait definition]

2008-10-09 Thread Adrian Chadd
On Fri, Oct 10, 2008, Adrian Chadd wrote:
 

 Now, you could go groveling through the kernel to try and figure out
 whether IOWAIT includes nfs activity (which I believe would be
 the disk IO related to NFS, but it could be other things I guess) and
 first see if doing userspace disk IO does a lot of IOWAIT. Grab bonnie
 or some other hard disk throughput testing thing, run it, and see how
 much time is spent in IOWAIT. If you've got almost no IOWAIT time when
 doing it locally but IOWAIT time when doing it over NFS, you could be
 right. If you have IOWAIT time on both, I'd poke the disk/controller/
 driver combo. 
 

I just re-re-read the OP and stuff and realise my assumptions don't match
your request. I assumed you wanted to know if IOWAIT included disk IO
time -and- nfs server time. But you're asking if IOWAIT include NFS
client time (ie, time spent on your box talking to the NFS server.)

Anyway, doing a bonnie++ or such test will still tell you. Run it locally
and run it over NFS. See if you get IOWAIT increases for both. That
should give you some hint as to whats going on. I've never seen IOWAIT
for NFS client traffic (ie, traffic from an NFS client talking to an NFS
server) but who knows, this is linux..



adrian

 2d,
 
 
 
 Adrian
 
 
 On Thu, Oct 09, 2008, David Kempe wrote:
  Grant Street wrote:
  
  
  I have a machine with a good proportion of IOWait 20-30%. It does have
  local disks and it performs operations on NFS mounts. I just wanted to
  be sure if IOWait includes NFS activity or not. I also want a way if it
  is NFS to be able to say for sure if it is a bottleneck on the nfs 
  client or server. NFS is not a linux machine so visibility is not 
  allways the best.
  
  dstat might help you correlate stuff.
  http://dag.wieers.com/home-made/dstat/
  
  dave
  -- 
  SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
  Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html
 
 -- 
 - Xenion - http://www.xenion.com.au/ - VPS Hosting - Commercial Squid Support 
 -
 - $25/pm entry-level VPSes w/ capped bandwidth charges available in WA -
 -- 
 SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
 Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html

-- 
- Xenion - http://www.xenion.com.au/ - VPS Hosting - Commercial Squid Support -
- $25/pm entry-level VPSes w/ capped bandwidth charges available in WA -
-- 
SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html


Re: [Fwd: Re: [SLUG] IOWait definition]

2008-10-09 Thread Ian Wienand
On Fri, Oct 10, 2008 at 05:50:52AM +0800, Adrian Chadd wrote:
 I've never seen IOWAIT for NFS client traffic (ie, traffic from an
 NFS client talking to an NFS server) but who knows, this is linux..

I would say this doesn't count to iowait either; see
fs/nfs/pagelist.c:nfs_wait_on_request() -- it appears to put itself
out of action without doing anything required to update iowait times
(see kernel/sched.c:io_schedule()).

I rekon Shehjar would know though... :)

-i
-- 
SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html


Re: [Fwd: Re: [SLUG] IOWait definition]

2008-10-08 Thread David Kempe

Grant Street wrote:



I have a machine with a good proportion of IOWait 20-30%. It does have
local disks and it performs operations on NFS mounts. I just wanted to
be sure if IOWait includes NFS activity or not. I also want a way if it
is NFS to be able to say for sure if it is a bottleneck on the nfs 
client or server. NFS is not a linux machine so visibility is not 
allways the best.



dstat might help you correlate stuff.
http://dag.wieers.com/home-made/dstat/

dave
--
SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html