Profile cover photo
Profile photo
Paul Turner
About
Paul's posts

Post has attachment
Modern metropolis.
Photo

Post has shared content
Yes, it has been 10 years since Lord of the Rings: Fellowship of the Ring came out.

Reunion photo: (by Sarah Dunn for Empire Magazine)
Photo

Post has shared content
When I was younger I wanted to be an Astronaut when I grew up.
One of the last pictures I took yesterday - Aurora Australis - the southern lights - dancing with the constellation Orion. An amazing moment.
Photo

Post has attachment
George R. R. Martin, no big deal.
Photo

I recently presented some CPU scheduling work we're evaluating to make
scheduling decisions (e.g. placement and wake-up) using global instead of
cpu-local state at LPC:
http://www.linuxplumbersconf.org/2011/ocw/proposals/309

One basic example of how this problem affects threads is as follows:
Suppose we have (equally weighted) threads {A,B} and {C} on CPU 1 and
2 respectively.

------A---------------B
-----C (by vruntime)

Since A,B,C all have equal weight they should all be receiving 1/3
(for a total of 2/3) of system resources. But, since we cannot
partition at a finer granularity than 1 CPU we in fact have {A,B}
receiving 0.5 < 0.66 and C receiving 1.0 > 0.66.

Yet, the current representation of time received is cpu-local, so,
when we do migrate B to C we currently consider it to be still ahead
of its fair share (since it was relatively ahead of A), resulting in
increased latency.

While the first half is an introduction of some CFS fundamentals;
there are some fairly interesting results on both accuracy and tail
latency tucked away at the end. *Caveat emptor; due to limited
testing so far (untrusted korg wreckage) there's a bunch of work still
to be done on making sure we're not hurting the /antagonists/ too
much!*

Slides are at:
https://docs.google.com/present/view?id=dgjvsb78_16g7d9g867
Wait while more posts are being loaded