2.0.28 and all kernels up to 2.1.14 are fine. I made a patch with all the
stuff I thought was relevant patch-2.1.15, and it was fine; it does not
look like it is a problem with any of the new checksumming or other
i86-tweaking code, at least not overtly. Bogomips stays constant at 50,
but dhry2reg drops from 70000/sec to 33000/sec with the problem kernels.
Most any other benchmark (plus normal interactive use) shows the problem
too.
If I make an extremely minimal 2.1.5 kernel with no serial devices, the IP
protocol (ONLY) and just the Ethernet driver (ISA Lance), I get the
problem. But if I take out IP support the problem goes away (even if I add
in IPX). If I have IP plus lots of other neat stuff and take away the
Lance driver, the problem also goes away. It is not based solely on kernel
size, as I can make rather large kernels with no problem and fairly small
ones that have the problem.
It sounds to me like something odd was done to loopback or Unix domain
code, which is impacted by having IP support together with a real Ethernet
device. Any ideas on this very squirrely problem? Anyone else seeing
unexpected slowness in >2.1.14 kernels?
I also heard back from someone else with an AMD DX4/100 processor who is
not experiencing this problem under 2.1.21. I'm not totally sure what if
any differences there are, but more data points would be appreciated
(either more AMD DX/4 users with no problems, or *anyone* who can also
reproduce it).
David
Niemi@linux.wauug.org 703-810-5538 Reston, Virginia, USA
------ Money talks, but it is wrong half of the time. -----