Firefly and PC GAMESS-related discussion club

Learn how to ask questions correctly

Alex Granovsky
gran@classic.chem.msu.su

Just want to share my rouble (ca. 3 cents as of now).

I believe the answer depends on the typical sets of jobs you
are planning to run. Some of them are not limited by I/O at all,
most of others can perform I/O asynchronously and that usually
hides I/O latency even on high-end workstations equipped by a
pretty standard server class SATA-II HDDs; and there are only
a few types of jobs that are I/O limited - the most notably case
is the conventional SCF. Note that Firefly typically adopts
"write once - read multiple" strategy for its working files
so that most of I/O is to read data, not to write them - and
this rather is good for SSDs... One of my friends ran lots of
conventional SCF jobs on the raid of two (rather old) SSDs
for more than one year in almost non-stop mode before one of
them died...

Sorry for late reply.

Regards,
Alex Granovsky



On Mon Jan 18 '10 7:46pm, Denis Zavelev wrote
---------------------------------------------
>Hello!

>After some previous discussions it appeared that SSD can probably be the best choice for disk subsystem of the workstation for Firefly. Really, what do we have: we can just buy single drive (or maybe 2 similar drives), make an appropriate number of filesystems on them and its performance will be enough in order not to be the limitation just for any computation. Of course, this solution is expensive (in comparison with SATA II), especially when you do need a lot of disk space for each thread. But besides the cost, we have the problem of reliability of SSDs. Oh yeah, manufacturer guarantees non-faulty work of SSD for 2 years, but I sipmply don't trust that those SSDs (2.5" for SATA II) that became available now are reliable enough. Yes, I won't loose the funds, but it's not a pleasure to visit the guarantee department every several months in order to exchange the faulty disk. IMHO, the age of SSDs will come 1-3 years later, not now. But we have to make the computations now!
>So I'm thinking about SAS HDDs (like Seagate Cheetah 300Gb).
>Of course, for best performance I'll have to get 4 such HDDs (1 HDD/core) and thus I'll have to get an additional SAS controller (intergrated SAS controller on some MBs can handle only 2 HDDs) so the first question: will it be OK to connect 2 HDDs to integrated controller and 2 other HDDs to PCI-E controller? Of course, it seems to be obvious that data transfer rate will be limited by HDD mechanics, but nevertheless I want to ask the experienced ones about such solution (as sometimes the solution that seems to be simple and obvious for non-expert produces the unexpected difficulties one after one). Of course, I understand that those HDDs must be separate (not in RAID) but does usual SAS controller (e.g. by LSI logic) support the parallel work of 2HDDs? Can there be any conflicts between onboard and PCI-E SAS controllers?