On Fri, 2003-05-16 at 11:43, Christopher Chan wrote:very interesting. this might have helped with a problem we ran into.
can you tell us what changes you made to the elevator?
the biggest change is a patch from Andrew Morton that will give priority to reads in the presence of heavy write loads; since reads in general are things the user waits on while writes are generally asynchronous, this is a very interesting optimisation
Does /proc/sys/vm/bdflush have anything to do with the journal on ext3?
We were seeing huge writes to disk at approximately the time when the journal would hit 60% full with default settings in /proc/sys/vm/bdflush. Those writes were causing time outs as the writes would go on for 15 or more seconds. (our journal is 128M in size and we are using full journal mode)
These writes were finally worked around by using:
1 200 0 0 500 2000 60 0 0
which meant we had largish disk writes for one or two seconds at frequent intervals but at least they are not causing time outs for our mysql clients.
In addition some constants used in the elevator about how agressive it can reorder have been tuned towards more interactive performance (at the cost of raw throughput/performance)
how can i adjust them? elvtune? any docs?
thnx
