I started Shambhala 0.6 at Cardiff.
The hit rate seems to have climbed a little, but I can't tell
if that is due to the reboot that went with the changeover. I'll
have a better feel for any improvements based on tomorrow's stats.
Everything seems to be working correctly.
-=-=
Another though on child process management..
Can we get any useful information from the time it would take the
parent to get past the "accept" mutex ?
e.g. if it takes 20 seconds for the parent to set the file lock, it's
pretty obvious that the hit rate is low, so a few unwanted children
can be zapped, whereas if the parent takes less than half a second,
then lots of children must have accept'ed in that time, and thus it
might be reasonable to spawn all the children up to the specified max.
This basically boils down to using the mutex as a means of measuring the
hit rate without having unwanted parent<->child communication.
I'd suggest that the parent lock and unlock multiple times, and the
average be used to estimate the current load.
Any thoughts ?. Does anyone care about this topic ?
rob
--
http://nqcd.lanl.gov/~hartill/
The hit rate seems to have climbed a little, but I can't tell
if that is due to the reboot that went with the changeover. I'll
have a better feel for any improvements based on tomorrow's stats.
Everything seems to be working correctly.
-=-=
Another though on child process management..
Can we get any useful information from the time it would take the
parent to get past the "accept" mutex ?
e.g. if it takes 20 seconds for the parent to set the file lock, it's
pretty obvious that the hit rate is low, so a few unwanted children
can be zapped, whereas if the parent takes less than half a second,
then lots of children must have accept'ed in that time, and thus it
might be reasonable to spawn all the children up to the specified max.
This basically boils down to using the mutex as a means of measuring the
hit rate without having unwanted parent<->child communication.
I'd suggest that the parent lock and unlock multiple times, and the
average be used to estimate the current load.
Any thoughts ?. Does anyone care about this topic ?
rob
--
http://nqcd.lanl.gov/~hartill/