Complete.Org: Mailing Lists: Archives: discussion: May 1999:
Re: [aclug-L] ntmag

Re: [aclug-L] ntmag

[Top] [All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index] [Thread Index]
To: aclug-L@xxxxxxxxxxxx
Subject: Re: [aclug-L] ntmag
From: Tom Hull <thull@xxxxxxxx>
Date: Tue, 04 May 1999 01:20:27 -0400
Reply-to: aclug-L@xxxxxxxxxxxx

I may be able to help on a couple points in the article, although
there is much that I am not totally clear on, either. I haven't
taken a close look at MP support in Linux 2.2, so I don't know
how fine-grained the locking is, or whether there is a LWP (light
weight process) concept and its associated interfaces. And I know
as little about NT as possible.

Some basic background: AT&T started working on UNIX MP support
for the SVR4 MP release, something like 10 years ago. The legacy
of that work shows up in UnixWare, Solaris, Irix, HP-UX, and
various other commercial Unix OS. Linux introduced very primitive
MP support in 2.0: there was a single lock around the entire
kernel. Linux 2.2 introduces more fine-grained locking (what
the article refers to as re-entrant kernel functions). However,
there are still several large subsystems in Linux that are
guarded by single locks: in particular, the file system and
the TCP/IP stack. (The article states that the lock is on the
read and write system calls. This may be correct: my information
is simply that ext2fs and tcp/ip are still single-threaded.)

The short story here is that commercial Unix has about an 8 year
head start in terms of MP support: all other factors being equal,
a commercial Unix will scale progressively better as more CPUs
are added. There's no real mystery to this. One major way to
improve MP performance is to replace big bottleneck locks with
finer-grained locking, so the CPUs run into each other less
frequently. Since each bottleneck you remove makes the next
most frequent bottleneck all the more visible, progress in MP
performance in incremental over time.

The same thing happens with other performance bottlenecks. Most
commercial Unix have support for asynch I/O, and many support
sendv (a system call which allows files to be copied straight
to a network socket without a sidetrip into user space; this is
what NT calls sendfile).

Regardless of how Linux 2.2 stacks up, I have no doubt that
some not-too-distant future Linux release will tackle all of
these issues. The only thing that has to happen is that Linux
hackers have to get their hands on larger MP equipment: this
is starting to happen, as evidenced by Penguin Computing's
announcement of a 8-Xeon Linux server.

I think the reasonable conclusion from all this is that if you
really need 4+ CPUs and really need to get every cycle out of
them, go with a commercial Unix system. On the other hand, if
you can get by with a little less performance, stick with Linux
(even in a 4-way MP). In the long run you'll be in better hands.

My impression is that, strictly in terms of performance, NT is
somewhere between commercial Unix and Linux. Microsoft has spent
a lot of money, both in building MP support and in benchmark
tuning, but they haven't come up with anything fundamentally
superior to Unix. My impression is that the emphasis in the
article on kernel threads is part "benchmark special" (some
finely tuned hack that favorably distorts the benchmark score
without providing any generally useful value), and part diversion
from the fact that Microsoft's process model is notoriously weak.
On the other hand, NT has so many negatives that I wouldn't
give any performance claims a second thought.

Quite frankly, I find that most benchmarks are so unrepresentative,
and that the tunings for them are so unorthodox, that I think it
would be foolish to give them more than a moment's thought. For
example, the current vogue in SpecWeb is to mount the file system
read-only (since SpecWeb only reads files) and to move the httpd
server into kernel space to avoid context switches (which is
possible because the benchmarks doesn't execute CGI scripts).
This is supposed to be an honest test of ISP performance?

As for TPC-C, check out:

The new champ, at 115395, is a 64-CPU Sparc, with 64GB memory and
1732 disks. The best NT score is down at 33935, a Compaq with
16-Xeon (4x4 cluster), with 13GB memory and 388 disks.

BTW, a good place to check up on the progress of Linux in
big-time Enterprise is:

phrostie wrote:
> a man i used to work with in austin is considering learning linux.
> his day job is as an nt administrator.
> today he sent me this article asking me what i thought .
> in truth most of what the author is talking about is over my head.
> my understanding is that linux has always had support for SMP,
> with 2.2 it is more efficient.
> i would like to give him some constructive feed back, but
> once again this is over my head.
> not too much nt bashing please :-).
> thankyou
> --
> phrostie@xxxxxxxxxxxxx
> Oh I've slipped the surly bonds of dos
> and danced the skies on LINUX silvered wings.

 *  Tom Hull  ::  thull@xxxxxxxx  ::

[Prev in Thread] Current Thread [Next in Thread]