[altq 12] Re: ALTQ internals

[Kenjiro Cho <kjc@csl.sony.co.jp>]

The two methods are identical but "avgidle" is scaled (x 32) in the
kernel for better precision and cheaper computation.

Let avgidle32 be (real) avgidle multiplied by 32.

	avgidle32 = avgidle x 32

Instead of the following equation

	avgidle = 31/32 * avgidle + 1/32 * idle

the kernel use

	avgidle32 = 31/32 * avgidle32 + idle	

		  = avgidle32 - 1/32 * avgidle32 + idle

Sally proposed this method in one of her notes on CBQ.
(Van Jacobson used a similar technique to estimate RTT in TCP.)

--Kenjiro

Fulvio Risso <F.Risso@cs.ucl.ac.uk> said:
>> Hello.
>> Checking the ALTQ internals, I discovered that the computation os the
>> *avgidle* parameter differs from the CBQ original proposal.

>> In fact,
>> ALTQ: altq_rmclass.c, function rmc_update_class_util
>>   avgidle+= idle - (avgidle >> RM_FILTER_GAIN)
>>   that means: avgidle = idle + (31/32) avgidle

>> CBQ (Floyd):
>>   avgidle= w * avgidle + (1-w) idle
>>   that means (if w= 31/32): avgidle= 1/16 idle + 31/32 avgidle

>> Why this difference?
>> Thanks a lot.

>> 	fulvio