Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!umich!yale!think!linus!alliant!neray
From: neray@Alliant.COM (Phil Neray)
Newsgroups: comp.sys.alliant
Subject: Re: non-zero ipl
Message-ID: <3723@alliant.Alliant.COM>
Date: 9 Mar 90 22:13:45 GMT
References: <1990Mar6.222236.7098@uokmax.uucp>
Organization: Alliant Computer Systems, Littleton, MA
Lines: 75

Reply-To: neray@alliant.Alliant.COM (Phil Neray)
Distribution:
Organization: Alliant Computer Systems, Littleton, MA
Keywords: Multi-Threading UNIX

In article <1990Mar7.013905.29368@Neon.Stanford.EDU>
weening@Gang-of-Four.Stanford.EDU (Joe Weening) writes:

>Alliant told us that Concentrix 5.5.1 would soon be sent to people
>having problems, so I debated whether to wait for it, but finally
>decided to roll back to 5.0.  This was at the end of January, and we
>haven't yet received 5.5.1, so I'm glad I decided to roll back.  We
>don't really need any of the features that 5.5 supports.

Concentrix 5.5 has some very useful features for systems that have
lots of users (and need high throughput), and/or systems with the
requirement for lots of aggregate I/O.

This is because Concentrix is now extensively multi-threaded.

Concentrix has always been a symmetric multiprocessing implementation
of UNIX (which means the kernel could run on any available
processor). With 5.5, the kernel itself has been parallelized,
allowing mutiple sections of the kernel to be executing simultaneously
on multiple processors.

Most super and mini-super multiprocessor UNIX implementations have
"coarse-grain" locking - ie., the entire file system gets locked, or
the entire virtual memory system, or the entire network system. 

Concentrix 5.5 has thousands of fine-grain locks that control access 
to critical data structures and code regions in the kernel. Also, 
Concentrix has some built-in strategies to minimize the duration of a
lock and the number of locks that can be held by a particular process.

(Other multi-threaded UNIX implementations that we know of include
Sequent and Encore.)

Multi-threading increases overall multi-user throughput, multi-stream
aggregate I/O, and multi-stream network throughput. For example, the
maximum number of simulated users supported on an FX/80, as measured
by the AIM Multiuser Benchmark, is now over 3x what it was with
Concentrix 5.0. (Of course, the FX/80 was designed for large number
crunching jobs, but most systems are also used for standard UNIX
development activities like compiling and file copying which is what
the AIM benchmark measures.)

The aggregate I/O with disk striping (TFS) is about 2x compared with
Concentrix 5.0, and the aggregate network I/O is also about 2x (this
is for smaller packet sizes which are representative of terminal and
X11 network traffic). "Lock-wait" has been virtually eliminated.

Note that multi-threading does NOT improve single-stream I/O (that's 
in development for a future release).

Also, this multi-threaded Concentrix has now been ported to the
FX/2800 (up to 28 Intel i860 40 Mhz 64-bit processors with 1 GB
memory), which should be a very high throughput system due to the
high scalar performance of the processors (>41 VAX MIPS, Dhrystone)
and increased I/O bandwidth (multiple 20 MB/s VME channels).

My apologies if this sounds too much like a commercial, but the
Concentrix developers (led by Jeff Jones and Tony Anzelmo) have done
some really fine work here and I thought you should know.

----------------------------------------------------------------------

Phil Neray			Domain:	neray@alliant.com
Alliant Computer Systems	UUCP:	{mit-eddie|linus}!alliant!neray
Littleton, MA 01460		Phone:	(508) 486-1429

-- 
Larry Fisher			Domain:	fisher@alliant.com
Alliant Computer Systems	UUCP:	{mit-eddie|linus}!alliant!fisher
Littleton, MA 01460		Phone:	(508) 486-1449