High performance computing on FreeBSD
ohartman at uni-mainz.de
Mon Feb 6 07:49:58 PST 2006
FreeBSd is now since 1996 my companion in scientific computing and
related server systems and also my favorite operating system for every
network stuff, firewalls and desktop systems I ever used.
Now going ahaed with 64Bit, FreeBSD 6.X has been canceled for desktop
systems due to the lack of a working JAVA in native 64Bit and especially
a working native 64 Bit OpenOffice environment.
Nevertheless, the experience of our group and especially of mine with
several flavours of Linux, used at our computer center and its network
performance and stability in comparison to FreeBSD's over the same time
period let me tend to ask for a FreeBSD based high performance computer
cluster more than such one founded on a Linux distribution. But there
are some open issues and those need to be discussed deeper.
First targets SMP/Node performance. I was very curious about SCHED_ULE
when introduced in FreeBSD 5.X and was said to deliver a performance
boost on SMP boxes. I'm still waiting for that to come true, every SMP
scaling benchmark that has been taken in our computer center said Linux
has the better SMP performance (on the same Opteron hardware, but I do
not have specific details about that, sorry).
Next point is the intercommunication of nodes. Infiniband or with
special Hypertransport coupplings nodes will be able to communicate very
fast. GBit LAN will be the least option, so the question is whether
plans for or ready solutions for the node connections are underway.
The last question refers to Fortran. Well, most of our scientists still
work with Fortran77 or Fortran90/95 and it is hard to bring them towards
C/C++, so the existence of good Fotran compiler will be essential. GCC
4.1/4.2 isn't standard in FreeBSD 6.X but many of other FreeBSD users
told me they use the port's gcc 4.X very successful. But I feel better
when the new GNU compiler collection will be the standard for FreeBSD.
This may sound weird for some of yours, but I like the ease of upgrading
software in FreeBSD which has reached a very, very high standard over
the past 10 years (and it isn't comparable to jarsh weirdness I
experienced with Linux, Solaris or Windows). So, utilizing standard
ports and the base compiler collection gives a very stable and high
quality platform - in my opinion.
All right, this above mentioned fundamentals should be the basis for a
small cluster system for numerical research.
I still looking for benchmark tests, pro and contra regarding BSD/Linux
(except the existence of better compiler software for Linux) and the
state of development of high performance node interconnect and
designated driver software.
Target hardware will be a four or six node Opteron/Athlon64 platform
with dual socket/dual core chips, with 4 or 8 GB local RAM and 200 GB
local SATA disk drives, but main disk array will be RAID system attached
via GBit LAN or, if possible, faster. The big question will remain in
how the nodes should be interconnected and what kind of OS will be able
to handle a specific interconnect (HTX/Infiniband).
In the case my questions are to unspecific or naiv, please excuse that.
More information about the freebsd-questions