[PATCH] staircase scheduler v6.4 for 2.6.7-rc3

[PATCH] staircase scheduler v6.4 for 2.6.7-rc3

[Con Kolivas]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Here is an updated full version of the staircase scheduler (not attached for 
brevity of email):

http://ck.kolivas.org/patches/2.6/2.6.7-rc3/patch-2.6.7-rc3-s6.4

25 files changed, 264 insertions(+), 545 deletions(-)

Changes:
Tiny bugfixes, cleanups.
Slight tweak to tasks waiting on I/O
Massive cleanups from wli - thanks!

Comments, testing welcome
Con
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFAxcvsZUg7+tp6mRURAgZFAJ4/iTCfBF/AotI70tbjPgW3xPy2TACggYgS
38J+HdFXv0lThCUzx07kNhM=
=eO0E
-----END PGP SIGNATURE-----

Re: [PATCH] staircase scheduler v6.4 for 2.6.7-rc3

[Peter Williams]

Con Kolivas wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> Here is an updated full version of the staircase scheduler (not attached for 
> brevity of email):
> 
> http://ck.kolivas.org/patches/2.6/2.6.7-rc3/patch-2.6.7-rc3-s6.4
> 
> 25 files changed, 264 insertions(+), 545 deletions(-)

There was no need to add the extra overhead of a flag to indicate that a 
task was queued for scheduling.  Testing whether run_list is empty 
achieves the same thing as reliably as the old array == NULL test did.

-- 
Dr Peter Williams                                pwil3058@bigpond.net.au

"Learning, n. The kind of ignorance distinguishing the studious."
  -- Ambrose Bierce

Re: [PATCH] staircase scheduler v6.4 for 2.6.7-rc3

[William Lee Irwin III]

On Wed, Jun 09, 2004 at 09:04:23AM +1000, Peter Williams wrote:
> There was no need to add the extra overhead of a flag to indicate that a 
> task was queued for scheduling.  Testing whether run_list is empty 
> achieves the same thing as reliably as the old array == NULL test did.

Overhead? Doubtful. Also, that requires the use of list_del_init()
while dequeueing, which is not in place now. Please do back the claim
with measurements. It should be easy enough to nop out set_task_queued(),
implement task_queued() via !list_empty(), and clear_task_queued() via
INIT_LIST_HEAD() for a quick performance comparison. But I'd say to
merge it even if there's no difference, as it's more self-contained.


-- wli

Re: [PATCH] staircase scheduler v6.4 for 2.6.7-rc3

[Peter Williams]

William Lee Irwin III wrote:
> On Wed, Jun 09, 2004 at 09:04:23AM +1000, Peter Williams wrote:
> 
>>There was no need to add the extra overhead of a flag to indicate that a 
>>task was queued for scheduling.  Testing whether run_list is empty 
>>achieves the same thing as reliably as the old array == NULL test did.
> 
> 
> Overhead? Doubtful. Also, that requires the use of list_del_init()

Yes, that's true.

> while dequeueing, which is not in place now. Please do back the claim
> with measurements. It should be easy enough to nop out set_task_queued(),
> implement task_queued() via !list_empty(), and clear_task_queued() via
> INIT_LIST_HEAD() for a quick performance comparison. But I'd say to
> merge it even if there's no difference, as it's more self-contained.
> 

Since the principle use of testing array for NULL or not was to find out 
if the task was on a run list it seems silly to have a flag to determine 
this.  All it does is provide an opportunity for the flag to not 
accurately reflect whether the task is really on a list or not.

It caused the number of files touched by the staircase patch to increase 
  by a factor of five which is another good reason to use the alternative.

Peter
-- 
Dr Peter Williams                                pwil3058@bigpond.net.au

"Learning, n. The kind of ignorance distinguishing the studious."
  -- Ambrose Bierce