Re: [RFC][PATCH] O(1) Entitlement Based Scheduler

[Helge Hafting <helgehaf@aitel.hist.no>]

Peter Williams wrote:
> Timothy Miller wrote:
> 
>> How about this:
>>
>> The kernel tracks CPU usage, time slice expiration, and numerous other 
>> statistics, and exports them to userspace through /proc or somesuch. 
>> Then a user-space daemon adjusts priority.
> 
> 
> Yes, the right statistics could allow these processes to be identified 
> reasonably accurately.  The programs in question would have the 
> following characteristics:
> 
> 1. low CPU usage rate, and
> 2. a very regular pattern of use i.e. the size of each CPU bursts would 
> be approximately constant as would the size of the intervals between 
> each burst.

There is no need for the regularity.  When I use a word processor, I
use it very irregularly.  Sometimes I type text, and wants each letter
typed to appear instantly.  This fits well with "low cpu usage" and
sudden short bursts.  There may be lots of long delays though while
I think about stuff to write.  So the intervals are irregular, I still
believe I should get the boosts as long as the bursts are small.
Doing something big (such as invoking latex on a big document)
is cpu-heavy, but then it is ok not to get the boost.
Current schedulers based on io-waiting gets this right already.
> 
> The appropriate statistic to identify the second of these would be 
> variance or (equivalently but more expensively) standard deviation. 
> Whether this problem is bad/important enough to warrant the extra 
> overhead of gathering these statistics is a moot point.  We had to 
> generate very high system loads on a single CPU system in order to cause 
> one or two skips in xmms over a period of a couple of minutes.

Well, perhaps you could give a slightly bigge boost to a very regular
thing like xmms.  But even that might have some snags, the load
might change a lot when doing midi in software, depending on how
many instruments are active simultaneously. There goes the
constant-size bursts.

Helge Hafting