One of the things I do on most of my systems is run the 
distributed.net RC5-64 key cracker, at nice +20, to soak up the space 
CPU cycles.

Some months ago, they released a new version of the client for 
NetBSD/alpha, and  the performance on my 500 MHz PC164 and 500 MHz 
DigitalServer 3305 (nee AlphaServer 800) doubled from 400 
kilokeys/sec to 800+ kilokeys/sec. Some clever person had obviously 
spent time optimizing the inner loop. Bravo for a factor of two.

However, I notice that when running the client, after some number of 
hours and key block iterations, performance of the client halves back 
down to 400 Kkeys/sec, and stays there. This happens on both 1.4.3 
and 1.5.1_ALPHA, with identical client code.

If I kill and restart the client, it delivers the 800 Kkeys/sec 
performance again, for a while...

Now, assuming that the distributed.net people haven't done anything 
wierd in the code and that the client is doing its best to run flat 
out with no waiting, what could NetBSD be doing that causes this 
performance drop?

Cache/TLB issues?

	curious,

	Erik <fair@clock.org>