Re: Cluster computing drawbacks


In article <87sly1wyyj.fsf@xxxxxxxxxxxxxxxxx>,
Ketil Malde <ketil+news@xxxxxxxxx> writes:
|> already5chosen@xxxxxxxxx writes:
|>
|> > [the trivial answer] "For the similar performance SMP reduces the
|> > cost of administration".
|>
|> You make it sound as if the only advantage is fewer boxes to stack and
|> cables to stumble in. However, there is a significant (at least in my
|> opinion) advantage for the user, in that a SMP system will run a lot
|> of software off the shelf with good performance, taking advantage of
|> SMP CPUs by spawning multiple processes and threads.

That is a Frequently Made Claim. However, it is very rarely justified
by evidence and, in most people's experience, is false.

There is actually very little off the shelf software that can get
much gain from SMP hardware by increasing the number of threads,
that isn't also possible by spawning the application onto multiple
boxes of a cluster. It does exist, but is far rarer than is
usually claimed.

Most of the third-party SMP-only applications don't scale beyond
a few CPUs, for reasons that we could go into but are irrelevant.
The fact is enough for this thread.

|> (Note that this is about SMP the user interface, not SMP as in a
|> certain type of CPU interconnect. SSI clusters may qualify.)

Server Side Includes?

|> This is probably heresy for the part of the HPC crowd who only want to
|> run Linpack really, really fast. But for many users of high (or at
|> least medium) performance computing, install and run, as opposed to
|> rewriting the application for MPI, makes a very real difference.

I know of very few such requirements. A few examples would make
me less doubtful of their rarity.

The point is that, if processes don't communicate, they can perfectly
well be run on separate boxes. And, with some provisos, this also
applies to ones that communicate through files. Similarly, anything
that can run effectively on a dual Opteron workstation isn't even
medium performance computing.

So you need to produce examples of requirements that can easily be
scaled to multiple CPUs using an inherently SMP method of communication,
but not so easily using an MPI one.


Regards,
Nick Maclaren.

.

Relevant Pages

  • Re: 16-Node Parallel System
    ... > We need a cost-effective 16-Node compute cluster for Academic ... Then you use any O/S kernel that is SMP aware, ... additional CPUs are added to the node). ... the interconnect of a SMP's 16 CPU interconnect will not ...
    (comp.parallel.mpi)
  • Re: VM fixes [4/4]
    ... >> If those old cpus really supported smp in linux, then fixing this bit is ... What I meant in this specific case being UP w/o preempt is enough to be ... So the only trouble here is SMP or PREEMPT. ... will fix the alpha arch with smp or preempt enabled on older cpus too. ...
    (Linux-Kernel)
  • Re: AMD Athlon 64 X2
    ... The CPU and motherboard also work with WSeB as well. ... Note - the motherboard I purchased came with BIOS rev 1013 and you ... have to flash the BIOS with a later version to get the CPUs ... I tried all of the recent SMP kernels but no luck. ...
    (comp.os.os2.setup.misc)
  • Re: signifanctly slowdown of FreeBSD 8.0-CURRENT/amd64
    ... although I haven't been doing any systematic benchmarking. ... What would be the penalties of running an SMP -CURRENT kernel on ... Or, counterintuitively, would an SMP kernel ... It seems my boxes, especially the ...
    (freebsd-current)
  • Re: AMD Athlon 64 X2
    ... Note - the motherboard I purchased came with BIOS rev 1013 and you have to flash the BIOS with a later version to get the CPUs recognized. ... I tried all of the recent SMP kernels but no luck. ... My expectation at present would be that most Athlon X2 systems will not work with OS/2 SMP and the ones that do are likely very unstable. ...
    (comp.os.os2.setup.misc)