Radix-tree pagecache for 2.5
Quoted 3 lines · Rik van Riel
Once you get over 'dbench 16' or so the whole thing basically becomes an excercise in how well the system can trigger task starvation in
get_request_wait.
It's neat you've identified that bottleneck.
dbench 192 also appears to trigger more swapin/swapout than usual with rmap based kernels about 12-15 minutes into the test; and it remains unusually high for the duration of the run. (not a huge amount of swapping, but vmstat 60 shows double digit numbers, rather than the more typical 0 with occasional single digits "spikes"). dbench 64 doesn't trigger this behavior on my test box.
I want diversity in the workloads. bonnie++ does a single thread, and tiobench is doing 1, 2, 4, and 8 threads. dbench fits in well at the other end of the spectrum.
The newest bench added to the lineup is OSDB on postgresql. If everything executes properly, 2.5.3-dj3 (which includes radix-tree) will win the first timer award for OSDB. :)
Signature · 1 line
Randy Hron
Quoted 6 lines · On Wed, 6 Feb 2002 [email protected] wrote:
Quoted 3 lines · Rik van Riel
Once you get over 'dbench 16' or so the whole thing basically becomes an excercise in how well the system can trigger task starvation in
get_request_wait.It's neat you've identified that bottleneck.
Umm, there's one thing you need to remember about these high dbench loads though.
They run fastest when you run each of the dbench forks sequentially and have the others stuck in get_request_wait.
This, of course, is completely unacceptable for real-world server scenarios, where all users of the server need to be serviced fairly.
regards,
Rik
Signature · 4 lines
"Linux holds advantages over the single-vendor commercial OS"
-- Microsoft's "Competing with Linux" document
http://www.surriel.com/ http://distro.conectiva.com/Quoted 2 lines
They run fastest when you run each of the dbench forks sequentially and have the others stuck in
get_request_wait.
One interesting part of tiotest is the latency measurements. Latency isn't printed by tiobench.pl though. I think it's valueable information (and wish I had it).
Quoted 3 lines
This, of course, is completely unacceptable for real-world server scenarios, where all users of the server need to be serviced fairly.
Agreed. I'm glad kernel hackers focus on latency too. :)
There are _some_ applications where throughput is critical though. I would prefer to measure both throughput and latency at the same time, but am not yet clear on how to deal with the Heisenberg principle.
Signature · 1 line
Randy Hron
Quoted 11 lines · Rik van Riel
Quoted 6 lines · On Wed, 6 Feb 2002 [email protected] wrote:
Quoted 3 lines · Rik van Riel
Once you get over 'dbench 16' or so the whole thing basically becomes an excercise in how well the system can trigger task starvation in
get_request_wait.It's neat you've identified that bottleneck.
Umm, there's one thing you need to remember about these high dbench loads though.
Quoted 6 lines
They run fastest when you run each of the dbench forks sequentially and have the others stuck in
get_request_wait.This, of course, is completely unacceptable for real-world server scenarios, where all users of the server need to be serviced fairly.
Right, and as we just discussed on irc, it's a useful effect - only if we control it, so that stopped or slowed processes do eventually get forcibly elevated in terms of IO priority so they can make progress, after being sat upon by more agressive/successful processes long enough. And we can control this, it's just going to take a few months to get basic issues of IO queues, RSS accounting, etc. out of the way so we can address it.
The trouble I have with paying a lot of attention to dbench results at this point is - we're measuring effects of kernel behaviour that is, at this point, uncontrolled and effectively random. IOW, we're not measuring the effects that we're interested in just now. If we need to know IO throughput, we need to use benches that test exactly that, and not other randomly interacting effects. As they said on Laugh-in many moons ago: 'very interesting, but useless'.
Signature · 1 line
Daniel
Quoted 11 lines · Rik van Riel
Quoted 6 lines · On Wed, 6 Feb 2002 [email protected] wrote:
Quoted 3 lines · Rik van Riel
Once you get over 'dbench 16' or so the whole thing basically becomes an excercise in how well the system can trigger task starvation in
get_request_wait.It's neat you've identified that bottleneck.
Umm, there's one thing you need to remember about these high dbench loads though.
Quoted 6 lines
They run fastest when you run each of the dbench forks sequentially and have the others stuck in
get_request_wait.This, of course, is completely unacceptable for real-world server scenarios, where all users of the server need to be serviced fairly.
Right, and as we just discussed on irc, it's a useful effect - only if we control it, so that stopped or slowed processes do eventually get forcibly elevated in terms of IO priority so they can make progress, after being sat upon by more agressive/successful processes long enough. And we can control this, it's just going to take a few months to get basic issues of IO queues, RSS accounting, etc. out of the way so we can address it.
The trouble I have with paying a lot of attention to dbench results at this point is - we're measuring effects of kernel behaviour that is, at this point, uncontrolled and effectively random. IOW, we're not measuring the effects that we're interested in just now. If we need to know IO throughput, we need to use benches that test exactly that, and not other randomly interacting effects. As they said on Laugh-in many moons ago: 'very interesting, but useless'.
Signature · 1 line
Daniel