Re[2]: [Devil-linux-develop] RTOS patch for DL

[Serge L. <se...@nn...>]

Dear Heiko,

Wednesday, July 27, 2005, 9:21:12 PM, you wrote:

>>>                 Without Perf patch    |  With Perf patch
>>>  max:                   93.7898s      |  21.2820s
>>> approx.  95 percentile: 61.2178s      |  18.1917s |
>>> Threads fairness:                     |
>>> distributin:            45.45/79.44   |  100.00/100.00
>>> execution:              42.06/78.00   |  99.62/99.93
>>>
>>
>> The most interesting result is "approx.  95 percentile:", i.e. the
>> most evident time of thread execute.

HZ> Are you sure you have the columns right? Right now it looks to me like
HZ> without the perf patch it's faster.
Yes, you are right. Without rtos patches kernel a bit faster, BUT
patched kernel more "threads fairness". The time of system reaction
greatly grows in the heavy loaded system, but with rtos patch time of
reaction is approximately constant. It's a quite logical - scheduler
with taking "priority" into consideration ( aka intellectual scheduler
) is slower as simple scheduler because it must check more conditions
before switch task.
But take a closer look on the test results - without patch max time of
thread execution is 93s, patched - only 21s. As I said in my previous 
message, the most interesting result is "approx.  95 percentile": with
the probability of 95% the time of test thread execution is 18s in the
patched system and 61s in the doesn't patched system. It's a serious
difference, I suppose, especially for applications with strong time
dependence (VoIP for a example).

Some information abt RTOS and these patches:
http://www.linuxdevices.com/articles/AT8906594941.html

-- 
Best regards,
 Serge                            mailto:se...@nn...