NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/comp/arch/8289

Internet Message Format | 1992-07-25 | 3.3 KB

Path: sparky!uunet!sun-barr!news2me.ebay.sun.com!seven-up.East.Sun.COM!uk-news.UK.Sun.COM!sunfen!adrianc
From: [e-mail] (Adrian Cockcroft - Sun UK - Cambridge SE - SPARC Specialist)
Newsgroups: comp.arch
Subject: Solaris (was: Sun 600MP Benchmark Anomaly)
Date: 24 Jul 1992 19:07:39 GMT
Organization: Sun Microsystems
Lines: 62
Distribution: world
Message-ID: <[e-mail]>
References: <[e-mail]>
Reply-To: [e-mail]
NNTP-Posting-Host: 129.156.129.1

In article [e-mail], [e-mail] (David Fuller) writes:
>
>In the case of Sun and their current OS, the critical discrimination between
>symmetrical and asymmetric is irrelevant since once you're out of a single
>CPU's ability to service OS requests you're out of business.  With a 2 CPU
>system that means system time is limited to 50%, which is reasonable.  With
>a 4 CPU system it's down to 25%, which may not be reasonable at high numbers
>of processes.
>
>So while the Sun offering is symmetrical it is not useful symmetry at this
>juncture.

If a system is running at more than 25% system time continuously I can
usually find plenty of scope for tuning to get it down to below 25%.
This has the desired effect in terms of getting a S600MP to work well
with 4.1.2.

>I'm sure that Solaris 2 will address this, it's just a question of
>where it degrades to a single thread and whether these places are harmful
>to system performance.

I have run a two processor S600MP with a Solaris 2 beta release with
both CPU's at over 95% system time (to show that it can be done). The
basic kernel, NFS code and the streams code at least are already multi
threaded and scale with CPU's.


The first version of Solaris 2.0 is now shipping (and is being used to
generate this posting :-) but only runs on sun4c (i.e. desktop
SPARCstation machines) since the S600MP arrived too late in the
development/test cycle. A second release at the end of this year will
support the S600MP, SuperSPARC, everything else and lots of
processors.

To oversimplify the issues and explain the essence of what Solaris 2.0 does:

There are two ways of adding MP to Unix. One is to place locks around
critical regions of code, which is the way that most small scale MP
implementations have been done in the past. The other, used in Solaris
2.0 is to place locks in data structures. This is much more scalable.

For example, if a section of the kernel messes with a table it could
have locks placed in the code so that only one CPU at a time can run
that code.  This is the way that a uniprocessor Unix kernel is written
using interrupt level masks to protect code regions. In Solaris 2 every
entry in the table would have a lock as part of its data structure and
every CPU can run the code simultaneously as long as they are modifying
different parts of the table.

One benefit is that the kernel is fully pre-emptible rather than having
preemption points and has a sub-millisecond guaranteed response time
for real time processes.

See ftp.uu.net:/systems/sun/papers/SunOS-5.0-MT for the full scoop.

Regards Adrian

---
Adrian Cockcroft -  [e-mail] or [e-mail]
Sun Microsystems, 306 Science Park, Milton Road, Cambridge CB4 4WG, UK
Phone [phone] - Fax [phone]    Sun Mailstop: ECBG01