* Re: [9fans] grid computing -- high performance?
2004-09-24 19:16 ` [9fans] grid computing -- high performance? Tim Newsham
@ 2004-09-24 19:25 ` Ronald G. Minnich
2004-09-24 19:45 ` Tim Newsham
2004-09-24 19:32 ` Eric Van Hensbergen
` (4 subsequent siblings)
5 siblings, 1 reply; 34+ messages in thread
From: Ronald G. Minnich @ 2004-09-24 19:25 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
On Fri, 24 Sep 2004, Tim Newsham wrote:
> Do people use the plan 9 compilers? It doesn't seem like the p9 compilers
> are suited for high performance number crunching (not that I think they
> are bad compilers).
And you know this ... how?
Although it would surprise me to find that 8c is as good as the Intel
compilers.
ron
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 19:25 ` Ronald G. Minnich
@ 2004-09-24 19:45 ` Tim Newsham
2004-09-24 20:39 ` Ronald G. Minnich
2004-09-24 21:38 ` geoff
0 siblings, 2 replies; 34+ messages in thread
From: Tim Newsham @ 2004-09-24 19:45 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
> > Do people use the plan 9 compilers? It doesn't seem like the p9 compilers
> > are suited for high performance number crunching (not that I think they
> > are bad compilers).
>
> And you know this ... how?
I don't, entirely. I'm partially asking about the quality of
the compiler code (perhaps indirectly).
The compiler implementation paper mentioned that the compilers emit
"medium quality" code. I assume that they were targetted towards their
purpose -- system code. Because of these factors, I assume (and that is
all it is, an assumption) that optimizations that would be important for
numerical computation but not very important for system programming didn't
receive much attention. Things like blocking, loop swapping, instruction
reordering, cache hints, combining operations into SIMD instructions, etc.
Is this not the case?
I'm not trying to imply that plan 9 is not a good system, or that
the compilers are poorly written. I'm just trying to learn more
about the system.
> ron
Tim N.
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 19:45 ` Tim Newsham
@ 2004-09-24 20:39 ` Ronald G. Minnich
2004-09-24 21:38 ` geoff
1 sibling, 0 replies; 34+ messages in thread
From: Ronald G. Minnich @ 2004-09-24 20:39 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
On Fri, 24 Sep 2004, Tim Newsham wrote:
> > And you know this ... how?
>
> I don't, entirely. I'm partially asking about the quality of
> the compiler code (perhaps indirectly).
Ah. You really need to take compiler performance app-by-app. We've found
on some codes that gcc does better than Intel 7.0 (!). It's just very hard
to draw general conclusions about how well code produced by a given
compiler will run.
ron
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 19:45 ` Tim Newsham
2004-09-24 20:39 ` Ronald G. Minnich
@ 2004-09-24 21:38 ` geoff
2004-09-25 2:54 ` boyd, rounin
1 sibling, 1 reply; 34+ messages in thread
From: geoff @ 2004-09-24 21:38 UTC (permalink / raw)
To: 9fans
Ken's description of the Plan 9 compilers, /sys/doc/compiler.ms,
describes in some detail how code is generated, and how the
work is split between compiler and loader (the assembler is
not involved in compilation).
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 21:38 ` geoff
@ 2004-09-25 2:54 ` boyd, rounin
2004-09-25 5:56 ` Bruce Ellis
0 siblings, 1 reply; 34+ messages in thread
From: boyd, rounin @ 2004-09-25 2:54 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
> Ken's description of the Plan 9 compilers, /sys/doc/compiler.ms,
> describes in some detail how code is generated, and how the
> work is split between compiler and loader (the assembler is
> not involved in compilation).
exactly. what a brilliant piece of work.
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-25 2:54 ` boyd, rounin
@ 2004-09-25 5:56 ` Bruce Ellis
2004-09-25 14:01 ` andrey mirtchovski
0 siblings, 1 reply; 34+ messages in thread
From: Bruce Ellis @ 2004-09-25 5:56 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
it's a very nice compiler to deal with. i added some x86
64 bit optims cause i needed them. it was entertaining and
educational. and then i got obsessed a bit and did some
multiply/divide optims. charles merged his stuff, jmk spent
far too much time testing and ... well release 2.1.10.2 is
what you get. only joking, the release is obviously the
mtime on sources.
brucee
boyd, rounin wrote:
>>Ken's description of the Plan 9 compilers, /sys/doc/compiler.ms,
>>describes in some detail how code is generated, and how the
>>work is split between compiler and loader (the assembler is
>>not involved in compilation).
>
> exactly. what a brilliant piece of work.
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 19:16 ` [9fans] grid computing -- high performance? Tim Newsham
2004-09-24 19:25 ` Ronald G. Minnich
@ 2004-09-24 19:32 ` Eric Van Hensbergen
2004-09-24 19:47 ` rog
2004-09-24 19:35 ` Christian Grothaus
` (3 subsequent siblings)
5 siblings, 1 reply; 34+ messages in thread
From: Eric Van Hensbergen @ 2004-09-24 19:32 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
On Fri, 24 Sep 2004 09:16:23 -1000 (HST), Tim Newsham <newsham@lava.net> wrote:
> I know that people experiment with plan 9 for grid computing,
> since the operating system is well suited for it. I was wondering
> if it is used for actual high performance work?
We may end up looking at that a little later this year - at least on
ppc64 power4-type systems.
>
> Do people use the plan 9 compilers? It doesn't seem like the p9 compilers
> are suited for high performance number crunching (not that I think they
> are bad compilers).
>
It all depends on what you want to do. We were trying to
hand-schedule some HPC code and the gcc compilers (and xlc compilers)
with their gratuitous cleverness completely screwed up our
hand-scheduling. If you really know what you are doing, simple
compilers are your friend.
> Have people ported optimizing compilers? If
> so, which ones, and are they available?
Hogan's GCC port might be something to look at (not that I consider
GCC to be an optimal compiler - but it'll have a different set of
optimizations than the Plan 9 compiler and might support stuff like
SSE).
-eric
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 19:32 ` Eric Van Hensbergen
@ 2004-09-24 19:47 ` rog
0 siblings, 0 replies; 34+ messages in thread
From: rog @ 2004-09-24 19:47 UTC (permalink / raw)
To: 9fans
it does depend a lot what task you're trying to compute.
in my recent experience, it's not unusual for an algorithm
to be available in binary-only format for a particular platform,
in which case a plan9-only solution would not be sufficient.
that's an advantage inferno has, i guess - it's straightforward to execute
the same binaries as everyone else, while still taking advantage
of the advantages of 9p2000/styx and a sane programming environment
for the actual grid infrastructure.
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 19:16 ` [9fans] grid computing -- high performance? Tim Newsham
2004-09-24 19:25 ` Ronald G. Minnich
2004-09-24 19:32 ` Eric Van Hensbergen
@ 2004-09-24 19:35 ` Christian Grothaus
2004-09-24 22:14 ` andrey mirtchovski
2004-09-24 20:23 ` jmk
` (2 subsequent siblings)
5 siblings, 1 reply; 34+ messages in thread
From: Christian Grothaus @ 2004-09-24 19:35 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
> Do people use the plan 9 compilers? It doesn't seem like the p9 compilers
> are suited for high performance number crunching (not that I think they
> are bad compilers). Have people ported optimizing compilers? If
> so, which ones, and are they available?
I use them for "number crunching" (in my case constructive
enumeration of graphs). The 8c compiler does some optimization,
even though it has no '-O' flag. I found that on x86, the speed is
comparable to that of 'gcc -O2'. The performance gain
with 'gcc -O[34]' is minimal in my case.
The compilation speed of 8c is unbeaten.
Christian
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 19:16 ` [9fans] grid computing -- high performance? Tim Newsham
` (2 preceding siblings ...)
2004-09-24 19:35 ` Christian Grothaus
@ 2004-09-24 20:23 ` jmk
2004-09-24 21:39 ` Tim Newsham
2004-09-24 22:13 ` andrey mirtchovski
2004-09-25 2:47 ` boyd, rounin
5 siblings, 1 reply; 34+ messages in thread
From: jmk @ 2004-09-24 20:23 UTC (permalink / raw)
To: 9fans
On Fri Sep 24 15:17:09 EDT 2004, newsham@lava.net wrote:
> ...
> Does the operating system have any support for SSE in the 386 port?
>
> On a slightly unrelated note, does 8c make any attempts to generate
> code that is reasonable for the pentium-4? For example, 16-byte aligning
> control flow targets?
>
> Tim N.
We have versions of 8a and 8l which have MMX/SSE support but they are not
installed because there is probably a compatibility issue with old object files.
No, there are no optimisations for the P4 (or PIII, PII, Pentium Pro, Pentium
or 486).
--jim
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 20:23 ` jmk
@ 2004-09-24 21:39 ` Tim Newsham
2004-09-24 22:14 ` jmk
0 siblings, 1 reply; 34+ messages in thread
From: Tim Newsham @ 2004-09-24 21:39 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
> We have versions of 8a and 8l which have MMX/SSE support but they are
> not installed because there is probably a compatibility issue with old
> object files.
For MMX, the p9 kernel doesnt need to do anything special, but the kernel
has to enable SSE and provide storage for the extra registers during
context switches, doesn't it? Is this in the kernel?
> --jim
Tim N.
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 21:39 ` Tim Newsham
@ 2004-09-24 22:14 ` jmk
0 siblings, 0 replies; 34+ messages in thread
From: jmk @ 2004-09-24 22:14 UTC (permalink / raw)
To: 9fans
On Fri Sep 24 18:07:47 EDT 2004, newsham@lava.net wrote:
> > We have versions of 8a and 8l which have MMX/SSE support but they are
> > not installed because there is probably a compatibility issue with old
> > object files.
>
> For MMX, the p9 kernel doesnt need to do anything special, but the kernel
> has to enable SSE and provide storage for the extra registers during
> context switches, doesn't it? Is this in the kernel?
>
> > --jim
>
> Tim N.
It's not in the distributed kernel, but that's where it needs to be, yes.
--jim
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 19:16 ` [9fans] grid computing -- high performance? Tim Newsham
` (3 preceding siblings ...)
2004-09-24 20:23 ` jmk
@ 2004-09-24 22:13 ` andrey mirtchovski
2004-09-24 22:52 ` Ronald G. Minnich
2004-09-25 2:47 ` boyd, rounin
5 siblings, 1 reply; 34+ messages in thread
From: andrey mirtchovski @ 2004-09-24 22:13 UTC (permalink / raw)
To: 9fans
> I know that people experiment with plan 9 for grid computing,
> since the operating system is well suited for it. I was wondering
> if it is used for actual high performance work?
>
It is a common misconception that Grid Computing == HPC, This is not
the case by a _very_ long shot. The problems Grid people are trying
to solve are of a much more general distributed nature -- how to get
computer A to utilize services available on computer B.
Plan 9's distributed design is what makes it so much better as a
platform than any grid middleware based on legacy UNIX and UNIX-like
OS's. Inferno solves the same problem somewhat differently, but it
gives you the ability to run on those legacy systems, which is the
hurdle Plan 9 seems to run most often against.
What you're looking for is an answer to the question 'how does Plan 9
do on clusters?'. Well, it hasn't been tested thoroughly on sizes
which most HPC people would find usable (the number of nodes Ron
frequently mentions) and the fact that 8c consistently generates
slower code than gcc by about 20% isn't much help either. Apocryphal
evidence suggests that it has been tried on 200+ CPUs on Tera systems,
perhaps others may elaborate on the results from that exercise.
The benefits you get from Plan 9 on the system administration side may
be compelling for some.
andrey
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 22:13 ` andrey mirtchovski
@ 2004-09-24 22:52 ` Ronald G. Minnich
2004-09-24 23:11 ` geoff
0 siblings, 1 reply; 34+ messages in thread
From: Ronald G. Minnich @ 2004-09-24 22:52 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
On Fri, 24 Sep 2004, andrey mirtchovski wrote:
> The benefits you get from Plan 9 on the system administration side may
> be compelling for some.
yes but ... Plan 9 is still harder to admin than the linuxbios/bproc
clusters we build here. It's way better than most cluster software systems
that run on Linux, but it's not the best you can get.
ron
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 22:52 ` Ronald G. Minnich
@ 2004-09-24 23:11 ` geoff
2004-09-27 13:51 ` Ronald G. Minnich
0 siblings, 1 reply; 34+ messages in thread
From: geoff @ 2004-09-24 23:11 UTC (permalink / raw)
To: 9fans
Maybe so, but aren't we arguing about values epsilon here?
(i.e., Plan 9 takes very little effort to administer, so
going from very little to very very little may not matter
much.)
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 23:11 ` geoff
@ 2004-09-27 13:51 ` Ronald G. Minnich
2004-09-27 18:45 ` geoff
0 siblings, 1 reply; 34+ messages in thread
From: Ronald G. Minnich @ 2004-09-27 13:51 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
On Fri, 24 Sep 2004 geoff@collyer.net wrote:
> Maybe so, but aren't we arguing about values epsilon here?
> (i.e., Plan 9 takes very little effort to administer, so
> going from very little to very very little may not matter
> much.)
no, because a small number times 1500 can start to look bigger than you
like.
ron
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-27 13:51 ` Ronald G. Minnich
@ 2004-09-27 18:45 ` geoff
2004-09-27 18:59 ` boyd, rounin
0 siblings, 1 reply; 34+ messages in thread
From: geoff @ 2004-09-27 18:45 UTC (permalink / raw)
To: 9fans
For many administrative things, you only have to do them
once per Plan 9 site (or cluster), not per machine. Is there
something specific you're thinking of that needs doing
for each machine?
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-27 18:45 ` geoff
@ 2004-09-27 18:59 ` boyd, rounin
0 siblings, 0 replies; 34+ messages in thread
From: boyd, rounin @ 2004-09-27 18:59 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
> Is there something specific you're thinking of that needs doing
> for each machine?
i think i agree with ron here. two or three machines is sorta fun,
but when it gets above 50 it gets tricky and beyond that ...
lunix is unmanageable once you get to 50 machines, unless you
a) like a total debacle or b) have spent a long time beating
them into shape.
lunix is unmanageable. the decentralisation that occurred
during the '80s was a colossal failure, but those boys
haven't learnt that yet.
plan 9 fixed all this by centralising the compute/disk resources
and moving back to the 'terminal model' [no disk, no fan] and
gluing it all together with a simple, connection based, coherent,
f/s protocol.
^ permalink raw reply [flat|nested] 34+ messages in thread
* Re: [9fans] grid computing -- high performance?
2004-09-24 19:16 ` [9fans] grid computing -- high performance? Tim Newsham
` (4 preceding siblings ...)
2004-09-24 22:13 ` andrey mirtchovski
@ 2004-09-25 2:47 ` boyd, rounin
5 siblings, 0 replies; 34+ messages in thread
From: boyd, rounin @ 2004-09-25 2:47 UTC (permalink / raw)
To: Fans of the OS Plan 9 from Bell Labs
From: "Tim Newsham" <newsham@lava.net>
>
> Do people use the plan 9 compilers?
no, we code in the instructions with the console switches.
dsw
^ permalink raw reply [flat|nested] 34+ messages in thread