that I should not have. It's not that I can't say no. They were good

patches but Alan Cox always maintained a list of outstanding issues so I

knew whether a patch was required or not. Without somebody keeping track

of these issues, it was kind of hard for me to say no.

There are still some issues with Linux, which are still

there, which make it not very enterprise scalable. Like the number of

processors it supports, SMP support, memory support, hardware support.

A lot of those issues are what 2.4 has been working very hard at

addressing so basically the scalability issues in an SMP environment are

gone. There's still the NUMA scalability to hundreds of CPU's and that's

going to be until the next stable release.

2.4 will scale to how many processors?

2.4 on realistic levels, I can say that we certainly tested 8-way and it

scales perfectly. It scales really well. The SpecWeb world record numbers

were done in an 8-way. It's interesting actually looking at the numbers.

You can go to the SpecWeb website and they have all the numbers for that

quarter and you can see Dell machines with 1 CPU, with 2, with 4 and with

8 and it's basically as perfect as it can get. I haven't seen the numbers

for 16 CPU machines but I suspect we'll do ok, at least for the web

serving kind of thing. But that's judging by how well we do beyond 8.

What is the upper memory limit for kernel 2.4?

It's basically set by hardware. On Intel it's 64GB. The SpecWeb

announcement was kind of fun because like any benchmark, they use

completely outrageous hardware. The record was done with 8 CPU, 64GB RAM

and 8 Gigabit ethernet cards. Support is really there. The issue for the

future is going to be clustering when you have non-uniform memory

accesses. That's the kind of work that SGI has been doing with their

IRIX'es. We're making sure that Linux clusters well on the Itanium.

What about the threads issues in Linux?

Linux threading is wonderful. The problem is that most people want to use

a standard threading model that was really designed for Solaris. It's

basically the Solaris Thread model that they've standardized and the

mapping from Linux threads is fairly ugly actually, and that's a problem

and we have to fix that. It involves extending the Linux thread model so

that it's easier to map from one to the other. That's actually one of the

things that's kind of on my wish list for 2.5 which is going to be the