Subject: performance oddities w/AMD 5x86 chips
To: None <port-i386@NetBSD.ORG>
From: Rafal Boni <rafal@mediaone.net>
List: port-i386
Date: 01/10/1998 01:09:31
Folks:
	I upgraded the CPU of my old 486/33 recently, to a AMD 5x86/133 (in
	the form of a Kington TurboChip-133) and am seeing some really
	strange performance data.

	The performance data comes from the distributed.net RC5-64 client
	(version v2.6403.315 in case anyone cares).  The same client runs
	on my P100 at a steady 130+K keys/s.  On the upgraded 486, I get
	very unusual data: at peak performance, I get up to 128K keys/s.
	Most of the time, however, the perfomance is a dismal 16K keys/s,
	almost an *order of magnitude* slower.  This seems unrelated to
	machine load, as when running this in single-user mode last night,
	I had the following string of results:

		128,000 kps, 126,000 kps, 116,000 kps, 110,000 kps,
		32,000 kps, 16,000 kps, 16,000kps...

	The kernel reports this as the CPU version/ID info:
		cpu0: family 4 model e step 4
		cpu0: AMD Am5x86 W/T 133/160 (486-class)

	Has anyone seen anything that would help explain this, or do people
	have ideas on where to start looking??  I assume the W/T in the CPU
	ID info indicated the CPU cache is write-through rather than write-
	back, but I don't know if that could be related... 

Thanks for any ideas!
--rafal

----
Rafal Boni                                                   rafal@mediaone.net