Printable View
Hello,
I read that IRIX is based on Unix System V. But i've been
all through nekochan forums, and still need answer to a
question:
What sets Irix apart from Unix? Besides the support for
Mips, what makes Irix different?
Any info or links would be really appreciated!
SL
In article <2rusi29q8tckfl54gbglvjhqtutgktr27m@4ax.com>,
Scaly Lizard <scalylizard@nospampleaseyho.com> wrote:[color=blue]
>Hello,[/color]
[color=blue]
>I read that IRIX is based on Unix System V. But i've been
>all through nekochan forums, and still need answer to a
>question:[/color]
[color=blue]
>What sets Irix apart from Unix? Besides the support for
>Mips, what makes Irix different?[/color]
[color=blue]
>Any info or links would be really appreciated![/color]
Primarily the tight integration of the SGI graphics hardware and
the multiprocessor/clustering support. The graphics performance
was way ahead of other vendors for years, and NUMALink (Non-
Uniform Memory Acces Link) allowed assembling large clusters of
varied pieces of hardware--all the systems did not have to be
identical.
Gary
--
Gary Heston [email]gheston@hiwaay.net[/email] [url]http://www.thebreastcancersite.com/[/url]
Astronomers have developed a definition of "planet" which excludes Pluto.
I'm developing a definition of "scientist" which excludes astronomers.
On Fri, 13 Oct 2006 00:42:03 -0000, [email]gheston@hiwaay.net[/email] (Gary Heston)
wrote:
[color=blue]
>In article <2rusi29q8tckfl54gbglvjhqtutgktr27m@4ax.com>,
>Scaly Lizard <scalylizard@nospampleaseyho.com> wrote:[color=green]
>>Hello,[/color]
>[color=green]
>>I read that IRIX is based on Unix System V. But i've been
>>all through nekochan forums, and still need answer to a
>>question:[/color]
>[color=green]
>>What sets Irix apart from Unix? Besides the support for
>>Mips, what makes Irix different?[/color]
>[color=green]
>>Any info or links would be really appreciated![/color]
>
>Primarily the tight integration of the SGI graphics hardware and
>the multiprocessor/clustering support. The graphics performance
>was way ahead of other vendors for years, and NUMALink (Non-
>Uniform Memory Acces Link) allowed assembling large clusters of
>varied pieces of hardware--all the systems did not have to be
>identical.
>
>
>Gary[/color]
Thanks, that's a big help!
Any other insights you might have would be great,
SL
Scaly Lizard wrote:
[color=blue]
> What sets Irix apart from Unix? Besides the support for
> Mips, what makes Irix different?[/color]
To add to what others wrote:
IRIX was perhaps the first to come out with
64 bit OS (running 64 bit pointer apps). And could also run
32 bit apps at full speed (no emulation here!).
[Hmm. No, I know I'm wrong as Cray was 64 bit for a long time.
Likely others I've simply forgotten.
Maybe the ability to run either/both ABIs on the same machine
at the same time without performance penalty is, I guess,
is the better point. The fact there
were *3* ABIs for a while is a historical oddity, not an advantage :-)
Maybe there's no point here at all :-)]
IRIX/MIPS was perhaps the only machine family spanning a wide
range of hardware were a (sensibly written) app could be run on all
machines
in the entire family and across multiple releases without
recompilation.[color=blue]
>From a little Indy to a large (for the period) multi-processor machine.[/color]
[I'm excluding the IRIX 6.3/6.4 period where there was an issue with
binary compatibility :-( The O2 (IRIX 6.3) folks were given far too
much freedom
to make changes so some apps would only run on an O2 unless recompiled
for IRIX 6.2 or IRIX 6.4.]
IRIX was perhaps the only OS that ran such a wide range of processor
counts
with a single-system-image. From 1 (Indy, O2) up to hundreds.
Where (as was said earlier) the processor cards could be different cpus
and
different cpu speeds. Not what we now call a 'cluster', as in a
cluster
there is a copy of the OS per processor. With IRIX one copy of the OS
ran the entire machine. By IRIX 6.5 up to thousands of processors
On 17 Oct 2006 14:27:21 -0700, "davea" <davea42@earthlink.net> wrote:
[color=blue]
>Scaly Lizard wrote:
>[color=green]
>> What sets Irix apart from Unix? Besides the support for
>> Mips, what makes Irix different?[/color]
>
>To add to what others wrote:
>
>IRIX was perhaps the first to come out with
>64 bit OS (running 64 bit pointer apps). And could also run
>32 bit apps at full speed (no emulation here!).
>[Hmm. No, I know I'm wrong as Cray was 64 bit for a long time.
>Likely others I've simply forgotten.
>Maybe the ability to run either/both ABIs on the same machine
>at the same time without performance penalty is, I guess,
>is the better point. The fact there
>were *3* ABIs for a while is a historical oddity, not an advantage :-)
>Maybe there's no point here at all :-)]
>
>
>IRIX/MIPS was perhaps the only machine family spanning a wide
>range of hardware were a (sensibly written) app could be run on all
>machines
>in the entire family and across multiple releases without
>recompilation.[color=green]
>>From a little Indy to a large (for the period) multi-processor machine.[/color]
>[I'm excluding the IRIX 6.3/6.4 period where there was an issue with
>binary compatibility :-( The O2 (IRIX 6.3) folks were given far too
>much freedom
>to make changes so some apps would only run on an O2 unless recompiled
>for IRIX 6.2 or IRIX 6.4.]
>
>IRIX was perhaps the only OS that ran such a wide range of processor
>counts
>with a single-system-image. From 1 (Indy, O2) up to hundreds.
>Where (as was said earlier) the processor cards could be different cpus
>and
>different cpu speeds. Not what we now call a 'cluster', as in a
>cluster
>there is a copy of the OS per processor. With IRIX one copy of the OS
>ran the entire machine. By IRIX 6.5 up to thousands of processors[/color]
Fantastic info, many thanks! I've been reading about NUMAlink since
Gary Heston's reply, and it's just amazing. 3.2 GB per second
bandwidth vs. gigabit ethernet's 100 MB/sec? Wow!
And Gigabit Ethernet has a 30 microsecond overhead for packets
as they are zipped and unzipped along the ethernet stack. NUMAlink
has gotten overhead down to ONE microsecond? Wow again!
Thanks again for your info, but wondering one thing that
i have not been able to find online:
I have not found a number i would like to know, just how
many procs can an Irix cluster run with one instance of
the OS? You say "thousands", but i wonder what is the
theoretical maximum number of nodes in an Irix NUMAlink
cluster?
SL
* davea:
[color=blue]
> IRIX/MIPS was perhaps the only machine family spanning a wide
> range of hardware were a (sensibly written) app could be run on all
> machines
> in the entire family and across multiple releases without
> recompilation.
> From a little Indy to a large (for the period) multi-processor machine.[/color]
Nope, MIPS/IRIX was not the only platform that could do that. Sun could
do the same with Solaris/SPARC as could HP with HP-UX on PA-RISC, or DEC
with Digital UNIX aka Tru64, or Siemens with SINIX aka ReliantUNIX, or
IBM with AIX on POWER.
All of these UNIX platforms offered binary compatibility, programs
running on a low end workstation could also run on big irons with dozens
of CPUs without modification or recompilation. That was pretty much
standard.
Benjamin
* Scaly Lizard:
[color=blue]
> Fantastic info, many thanks! I've been reading about NUMAlink since
> Gary Heston's reply, and it's just amazing. 3.2 GB per second
> bandwidth vs. gigabit ethernet's 100 MB/sec? Wow![/color]
NUMAlink does up to 6.4GB/s.
[color=blue]
> And Gigabit Ethernet has a 30 microsecond overhead for packets
> as they are zipped and unzipped along the ethernet stack. NUMAlink
> has gotten overhead down to ONE microsecond? Wow again![/color]
You are completely ignoring that NUMAlink and Ethernet have totally
different purposes. NUMAlink is a high bandwidth low latency connection
for coupling computer nodes together. Ethernet on the other side is a
extremely flexible standard for connecting lots of devices together in
big networks with means of redundancy easily. Comparing Ethernet and
NUMAlink is like comparing apples and oranges.
[color=blue]
> I have not found a number i would like to know, just how
> many procs can an Irix cluster run with one instance of
> the OS? You say "thousands", but i wonder what is the
> theoretical maximum number of nodes in an Irix NUMAlink
> cluster?[/color]
IIRC the limit is 1024 CPUs. IRIX strengths is that it allows a huge
single system computer. But today more and more HPC tasks are done by
cheaper clusters.
Benjamin
Benjamin Gawert wrote:[color=blue]
> Nope, MIPS/IRIX was not the only platform that could do that. Sun could
> do the same with Solaris/SPARC as could HP with HP-UX on PA-RISC, or DEC
> with Digital UNIX aka Tru64, or Siemens with SINIX aka ReliantUNIX, or
> IBM with AIX on POWER.
>
> All of these UNIX platforms offered binary compatibility, programs
> running on a low end workstation could also run on big irons with dozens
> of CPUs without modification or recompilation. That was pretty much
> standard.[/color]
+ DEC/Compaq/HP - Open VMS on VAX/Mips/Alpha (not Unix OS)
> I have not found a number i would like to know, just how[color=blue]
> many procs can an Irix cluster run with one instance of
> the OS? You say "thousands", but i wonder what is the
> theoretical maximum number of nodes in an Irix NUMAlink
> cluster?[/color]
Well who knows, but we are probably limited by few factors (interconnect
fabric bandwidth, cache coherency - cacheline updates, directory cache
size). We have working instalation of Origin 2xxx up to 2048 CPUs, O3xxx
up to 512 CPUs as i know, but i might be wrong as i am not working with
SGI's anymore.
It is probably not so wise to put more CPUs in single system, as we
would probably saturate whole design and decrease performance, also
there is problem with power/cooling factors.
On the other hand it is possible to build 4096 CPUs system, but design
should be revisited.
Origin architecture uses cache coherent NUMA, that means all coherency
is done entirely in hardware so we don't need any special tricks in OS,
but IRIX is NUMA aware and could use different approaches to minimaze
latency and how to deal with process affinity.
* Silvo Bozovicar:
[color=blue]
> + DEC/Compaq/HP - Open VMS on VAX/Mips/Alpha (not Unix OS)[/color]
Exactly.
Benjamin
In article <sm9bj2t1acv0931eia12joj3fhl20vvej4@4ax.com>,
Scaly Lizard <scalylizard@nospampleaseyho.com> wrote:
[ ... ][color=blue]
>Fantastic info, many thanks! I've been reading about NUMAlink since
>Gary Heston's reply, and it's just amazing. 3.2 GB per second
>bandwidth vs. gigabit ethernet's 100 MB/sec? Wow![/color]
[ ... ]
That bandwidth does come at a price; a few years ago a customer bought
an Origin 300 system consisting of two chassis ("C-bricks") that were
NUMAlinked together with a 1 meter cable--that cable cost $5000.
That server also had a dedicated 20A 230V outlet to feed the 3000VA
UPS in the bottom of the rack. I well remember installing that UPS.
By myself...all 125 pounds of it.
Gary
--
Gary Heston [email]gheston@hiwaay.net[/email] [url]http://www.thebreastcancersite.com/[/url]
Astronomers have developed a definition of "planet" which excludes Pluto.
I'm developing a definition of "scientist" which excludes astronomers.
Silvo Bozovicar wrote:[color=blue]
> Benjamin Gawert wrote:[color=green]
>> Nope, MIPS/IRIX was not the only platform that could do that. Sun
>> could do the same with Solaris/SPARC as could HP with HP-UX on
>> PA-RISC, or DEC with Digital UNIX aka Tru64, or Siemens with SINIX aka
>> ReliantUNIX, or IBM with AIX on POWER.
>>
>> All of these UNIX platforms offered binary compatibility, programs
>> running on a low end workstation could also run on big irons with
>> dozens of CPUs without modification or recompilation. That was pretty
>> much standard.[/color]
>
> + DEC/Compaq/HP - Open VMS on VAX/Mips/Alpha (not Unix OS)[/color]
But Ultrix/OSF-1/Tru64 *was/is* a UNIX OS. And, IIRC, Ultrix ran on
MIPS, where OSF-1/Tru64 was Alpha only (but I could be wrong).
Cheers,
Gary B-)