Hello Managers !
My last question on switched and shared FC RAID arrays
(see original post at the end) triggered a lot of people
to answer (a preliminary warm thanks to you all !). As the
picture is not clear yet, I'll try to clarify what we intend
to do and will post a worked out, extended summary when the
information gets consistent. A couple of people wrote that
they are also highly interested having a clear view of that
complex field. As Friday (the deadline for our order for the
RAID system) is coming near, I hope that some experts have
definitive answers.
This is what we INTEND to do:
5 TrueUnix 5.0A ES40 Servers and 3 Sun Solaris 7 (2 E450 and 1 E4500)
no MemoryChannel/no TrueCluster no clustering
one FC-PCI adapter per host
--T1---T2---T3---T4---T5--Switched--S1---S2---S3--> Users
+ + + + + 100M LAN + + + workstations
+ + + + + + + +
+ + + + + + + +
------------------------------------------------
= 16 - Port FC switch Fabric =
------------------------------------------------
+ + + + + +
+ + + + + +
--------------------
= RAID 3 x HSG80 = No redundancy
= EMA 12000 D14 = (not mission-
Only data files, =- - - - - - - - - = critical)
no swap, no OS = =
system files, = True64 AdvFs = RAID 5 or 3/5
no /home = several filesets =
=- - - - - - - - - = 36 GB 1" or 72 GB 1.6"
= Solaris 7 UFS = 14 or 10 disks/shelf
-------------------- 9 shelfes total
The 36 GB 1" or 72GB 1.6" hard disks shall be available soon.
The True64 and Solaris Units shall be completely separated,
necessary exchanges will be done via NFS (do we need to
dedicate a complete HSG80 to Solaris or does it suffice to
separate the Units ?)
On the ES40s, 20 to 40 disjoint SMP processes are number
crunching satellite data (Envisat/Mipas) which is read at the
beginning of each process (~350 Mbyte) and the results
(~5-50 Mbyte) written back at the end, possibly some state
writing to logging files in between. The runtime of each
process is between 10 Minutes and several hours. The input
amount is about 350 Mbyte per process, it is likely that several
processes READ the same data files from the RAID array at the
same time. It is also likely that several processes WRITE to
DIFFERENT files in the same directory of a unit at the same
time.
On the Suns, two different database programs (Oracle and an
object-oriented one) maintain the metadata of the files
(essentially copies of the data file headers), serving as a
backend to a graphical user interface.
The processes will be managed by a job scheduler (self made)
which partially uses LSF.
THE PROBLEM:
Apart from the hot topic of using two OSes, the one most argued on
is if the ES40s can work concurrently on the same AdvFs filesystem
(without actually writing to the same file, but possibly reading the
same one) without clustering. Most people suspect that this will result
in a corrupt system because filesystem caches of the different hosts
become inconsistent. This apparently can only be overcome by clustering
the machines, so they can coordinate their accesses. But we don't
need a cluster for anything else than this, so it will be wasted
both money and performance. Having the same filesystems presented
to all hosts (of one OS) is essential to our mission. So what
we really need is the functionality of a host having mounted the
filesets exporting it to all other hosts (which, as a real solution,
has been abandoned after throughput analysis), but the controller
seems not to serve it in this way. A Network Attached Storage (NAS)
would do it the right way, but only over LAN with a very poor
performance (estimations showed that the connection must have
about 4 times Gigabit ethernet throughput to get it near the
solution proposed). So it seems that SAN is currently not as universal
as many people expect it to be (and what would be desirable to have).
By the way, do we need also a cluster service on the Suns for the same
reason ?
Is there any way to get this configuration into work ? Or must we
waste our sparse money resources buying MemoryChannel and TrueCluster
Licenses ?
--
Dr. Udo Grabowski email: udo.grabowski_at_imk.fzk.de
Institut f. Meteorologie und Klimaforschung II, Forschungszentrum Karslruhe
Postfach 3640, D-76021 Karlsruhe, Germany Tel: (+49) 7247 82-6026
http://www.fzk.de/imk/imk2/ame/grabowski/ Fax: " -6141
Received on Wed Jun 28 2000 - 09:52:00 NZST