PC GAMESS/Firefly-related discussion club

Learn how to ask questions correctly

Alex Granovsky
gran@classic.chem.msu.su

Yes we do have SSD drive, a 120GB OCZ Vertex OCZSSD2-1VTX120G
(Mac Edition) one. For testing purposes, we used Core i7-based system
described here the only difference is that OS (note this is Win2K8 R1, not R2)
was upgraded to SP2 and all the latest updates from MS were installed;
and the SSD drive was added as the fifth SATA II disk on ICH10R.

According to our tests, the sustained read speed is ca. 240 MB/s,
while write speed is ca. 150 MB/s. These numbers are close to that
of device's specification.

The performance is already really good with Firefly. The only thing
that could improve it further is the support of asynchronous direct
unbuffered sequential writes - and we plan to add this feature to the
Firefly's I/O library soon enough to get the most of any SSD under
any OS. At the same time, the I/O library will be extended to allow
files to transparently span/stripe several physical disks.

MP2 gradient code is indeed very interesting case to test,
especially because all the details of its I/O can be changed
and/or tuned via input file variables. We used our Test 7
MP2 gradient code benchmark for testing.

We performed two parallel runs using HDDs (the first one using four
HDDs, one per each Firefly's instance; the second one - with a great
deal of reluctance - using single HDD for all four processes). Next,
we performed several parallel runs using single SSD (filesystem:
freshly formatted NTFS, cluster size 4 kb) while changing various I/O-
related options to find the optimal ones.

What was actually varied between runs are the following options:

 $SYSTEM mxiob(1)=

 $mp2 ioflgs(1)= 

 and 

 $mp2grd nbufs= rsize1= rsize2=

The first option is the Firefly-wide maximum size of I/O request
for read and write operations (in kilobytes). The current defaults
are mxiob(1)=128,128

The second one can be used to turn on direct unbuffered read and
write operations for random I/O for DASORT and DAFL30 files. This
is the recommended option under Win2K8 and Vista, see MP2 gradient
page for more information.

Finally, the third set allows one to modify number of I/O buffers
used by MP2 gradient code and record sizes for files DASORT and
DAFL30. Some more information on these options is available on
the MP2 gradient page.

The results of our tests are attached - so everyone can decide
for himself. Any comments and questions are very welcome.

Regards,
Alex Granovsky




On Wed Nov 18 '09 0:15am, Slawomir Janicki wrote
------------------------------------------------
>I don't have first-hand experience yet, but I am planing to use two SSD drives in my next i7 PC. From my research it seems that the choice of the drive is critical to long term performance. An article in AnandTech gives a good overview of SSDs: http://anandtech.com/storage/showdoc.aspx?i=3631http://anandtech.com/storage/showdoc.aspx?i=3631
>There are two issues to consider:

>- SSD drives read in 4kb blocks, but write in 512kb blocks. The controller has to manage this, but careful adjustment of the record size in Firefly will help for sure.
>- the MLC NAND gates can fail after about 10,000 writes, the SLC NAND gates can fail after about 100,000 writes. Again, the controller has to manage the lifespan of the drive.

>It seems that the the SLC drives from Intel X-25E have the upper hand over Intel X-25M (second generation) MLC drives, which are better than most of the other MLC drives. However, they can still loose 15f performance after some use. Under Windows 7/2008 R2 these drives can use the TRIM instruction to "defragment" themselves and gain back most of this loss. However under XP/Vista/2008 R1 the the dives with the Indilinx need a command prompt utility which may or may not be reliable.

>There are new drives coming to market with better prices than Intel. The Solid 2 series is built based on the same flash as Intel X-25M.

>The best buys seem to be the Intel X-25M and OCZ Vertex II. The "do not buy" label belongs to Samsung, Western Digital, and anything OEM. Which means, don't upgrade your Dell, Apple, Lenovo, or HP to SSD, you are likely to get Samsung junk.

>The good drives saturate the 3Gb SATA bus on sequential read, so I am thinking that each drive can service four cores of i7 920 or 860. I am sure that getting the top performance will need tweaking with the record sizes in Firefly.

>Slawomir
>
>
>On Tue Nov 17 '09 6:57pm, Thomas Patko wrote
>--------------------------------------------
>>Dear PC GAMESS / Firefly Users:

>>Does anyone have any first hand experience using any of the newer solid state disks (SSD) as temporary directories for I/O intensive jobs such as MP2?  While it is generally recommended to only use one traditional HDD for each temporary directory (per core) I wonder if the SSD can support multiple temporary directories for disk intensive I/O jobs as they should perform better for simultaneous read/write operations.  If so, how many temporary directories (cores) can one SSD handle simultaneously without significant slowdown (like traditional HDD)?

>>Any manufacturer recommendations and timing reports would be appreciated.  While the SSD prices have come out of the stratosphere it seems that the SSD are still rather expensive so I am weighing the cost options, although space becomes a consideration for machines with very many cores.

>>I am also interested in any first hand / anecdotal accounts about MTBF (other than what the manufacturer report for general use versus this significantly more severe service Firefly use).

>>Cheers,

>>Thomas

This message contains the 270 kb attachment [ mp2grad_ssd_vs_hdd.zip ] MP2 gradient benchmarks on SSD vs HDD