2010年8月 7日

[ANNOUNCE] BFS CPU scheduler v0.323 for 2.6.35

This is to announce the availability of the updated BFS CPU scheduler for
linux kernel v2.6.35.


http://ck.kolivas.org/patches/bfs/2.6.35-sched-bfs-323.patch


General BFS documentation:


http://ck.kolivas.org/patches/bfs/sched-BFS.txt


This will be included in 2.6.35-ck1 which is to be announced shortly. If time
permits and demand is present I will slowly port v323 to some of the earlier
kernels again.


Changes since the last announced version (0.318):

The most significant change has been some architectural change to work with
changes in the mainline kernel as of 2.6.35. As suspend / halt and cpu offline
code has been changed, the whole offline cpu/online cpu code had to be
modified in BFS to suit. Previously an affined task that was bound to an
offlining CPU had a temporary affinity placed in an "unplugged" CPU mask. This
system was fragile and placed an extra cpumask_t into task struct, and really
only worked for offlining of all CPUs during suspend/halt when all came back
online. Now, tasks that are affined to CPUs that are not currently online can
temporarily run on any CPU. This makes the system more robust and should work
properly on all types of suspend/halt. Thanks to Radoslaw and others in #ck
for pointing out this issue and helping test the modified version. None of
this should be visible in userland unless you were already having suspend
issues.

Alexei Podtelezhnikov pointed out that the code that modified the rr_interval
was ugly and helped rework it to do close to what I desire in a much nicer
fashion. Once again, the rr intervals have now been shrunk a little bit
further, and do not go above 24ms on any sized machine by default. They can
still be modified via /proc as per always. This may decrease throughput
slightly but should keep latencies much more stable on the many-core machine.

Also thanks to Alexei, a number of microoptimisations were done in the
SCHED_ISO refractory testing code. Should not be user visible.

Some debugging checks that weren't relevant to BFS were removed, and
unnecessary preempt disable/enable calls were removed. May be user visible if
you were getting warnings in dmesg previously.

CPU load calculation for use by the cpu frequency subsystem was improved to
properly tell how busy each individual cpu was. This may cause noticeable
improvements in how quickly cpufreq adapts to load in the ondemand governor.

The nohz_ratelimit function new in 2.6.35, and already noted to be buggy and
slated for removal in 2.6.35.1 was added as a no op to prevent this bug from
showing up on BFS, and to make easy patching come 2.6.35.1

Random other minor cleanups that I can't remember.


Enjoy!

--
-ck
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


投稿者 xml-rpc : 2010年8月 7日 09:07
役に立ちました?:
過去のフィードバック 平均:(0) 総合:(0) 投票回数:(0)
本記事へのTrackback: http://hoop.euqset.org/blog/mt-tb2006.cgi/97507
トラックバック
コメント
コメントする




画像の中に見える文字を入力してください。