Dear all,
Recently the technology in SDD disk have been developed by increasing the read-write velocities and the "lifetime" of writes such hundreds of TB. Even so, the available memory is very low and expensive.
On the other hand, CCSD(T) calculations easily can request more RAM than cluster have, requesting then read-write process in the disk, being a disastrous bottleneck in the time-cost.
For this reason, my question is: If we employ a SDD disk in CCSD(T) calculations in a write-read process, which will be the lifetime expected to that disk? What are the unit of "write-read" process?
For example, with a CCSD(T) that will require 200GB in disk (without taking into account RAM section), what will be the impact on the SDD? It will reduce only 200GB of "life" or will be a more complex cost?
Thank you very much for you time.
I am not an expert, but it is generally acknowledged that the lifetime of a SSD is affected by the use of large files, since there is not a defragmentation mechanisms. For my experience, Gaussian (and possibly different computational codes as well) saves very large files (chk and ref files mainly for Gaussian) that would lead the SSD to become progressively slower. In addition, there would be the risk to get close to the capacity of the disk. For these reasons, although TRIM may help, at the moment I wouldn't adopt a SSD in a computing workstation.
Dear Massimiliano,
Thank you very much for your answer. In reference for the files that gaussian uses in temp or scratch, this type of file are being write and/or overwrite continuously? If it is true, then, gaussian can write "in total" an important hundreds of GB for only one calculation. Then, as you explain, the calculation will deteriorate the SDD. Even so, if this temp files are statically written (no overwrite) and they are equivalent to the disk usage given, i.e 200GB, then the cost could be acceptable for a very specific calculations. Of course, for a cluster use with hundred of calculations per month it is impossible to implement.
Best regards,
Joaquim
Hello, I recently realized that the transfer rate to four sas/sata3 15K rpm disks configured in stripping (raid 0) can be equal or even exceed the transfer rate to a single ssd.
The transfer rate was 400MB/s and it is more than the transfer to ssd by sata2 interface. In conclusion to your expectations: Please try raid 0 configuration with 4 ssd or 15k rpm disks, but your raid controller should be able of top speed. I wish you success, Jerzy
In Gaussian, the files are continually overwritten. So, the SSD will suffer a lot from even a single calculation. Anyway, the besr bet is, for me, to spend money to buy more RAM.
Dear Massimiliano, Jerzy, Krysztof and Luc-Henri,
Thank you very much for your kind and very informative answers. At this point I arrive to the conclusion that SSD only can be applied to a cluster at extremely expensive way and with an important on the restriction in the memory ( for those calculation with a huge memory requirement).
On the other hand, it must be notice that computer technology it has been advancing a lot in the last few years (after 4 years I will change my PC and, for the same price I will double the capacities).
Even so, the adaptation of the cluster services to the new technologies it seems more difficult and very expensive, such the implementation of CUDA-GPU hardware, clusters with higher capacities for RAM and faster frequencies, and no less important, the new technologies in motherboards (with more slots and RAM capabilities, PCI and faster connections).
Finally, I would like to remark a very interesting paper that links in this Question Forum:
Article Computational Chemistry: The Fate of Current Methods and Fut...
Best regards,
Joaquim Rius
Note that ${PROGRAM_THAT_WAS_MENTIONED_ABOVE} is far from optimal for large coupled-cluster runs. You might want to also look into other programs, such as PySCF, Psi4 or CFOUR, which are all freely available, the first two being free software and the last free for academic use.
This is a very interesting topic for me. I have bought two servers with the same specs, but one has two SAS HDDs, and the other one two PCI SSDs.
Regarding the performance, I can comment that Gaussian16 is running approx. 20% faster, and that the writing in the drive corresponds to the .chk and log files, and of course the scratch (which is about 20-90 GB per calculation).
Dear Andra,
This is very interesting then. Please, could you give feedback with the impact in the shelf life of the SDS by their intensive use?
Best regards,
Joaquim
- Badges
- Science topic