[mythtv-users] Source of my IOBOUND problem?

David Raine rainecc at gmail.com
Thu Jun 7 18:49:00 UTC 2007


> I am seeing 40K sectors/second of activity with the system maxed out.  This
> converts to 20MB/s.  My HD is rated at 100MB/s as a PATA drive.  No way I

Running hdparm on a PATA drive on a supermicro P4 server mobo (the
only one I could find) gives me 15MBytes per sec throughput for
buffered disk reads. I think your disk is maxed due to chipset bus
constraints. The signs are there - you are running mythtranscode at
the same time (lots of reads and writes), iostat is showing circa 45%
iowait while the disk utlisation is almost 100%, vmstat matches these
numbers and also suggests a small amount of memory constraint (so
!=0), vmstat shows total io block transfers of >16k at peak - these
are 4k blocks, for a total of 64MB/s sadly not split by disk. The
numbers are way over max bandwidth for a single PATA drive doing this
lot, and probably 50+% is reading from the capture card, even so -
that's a lot.

There may be some mileage in disabling access time updates (noatime
option in ext3),  or using a different filesystem  but it won't give
you much and should only be done on a dedicated media partition
anyway.

The number of context switches and the system time will be related and
these are probably a symptom of the iowait; Remember linux will start
to spend more and more time in the overhead of servicing the
interrupts for the io, buffers will start to fill, run queue will
increase and overall io will decrease...It's a kind of negative
feedback that can only be delayed by tuning if more memory is
available. While that works for small bursts, your overall throughput
numbers are higher than a single disk channel can maintain. Time for
another disk.

[/random meanderings]


More information about the mythtv-users mailing list