The RTDC
Processing SMA Data
1.2 m Telescopes
AST/RO
Extra

Comparing Disk Speeds on Hydra


Disk options
Hydra offers its users three tiers of disks from which to run their jobs.

TypeDescriptionLocation
NFSNetApp disks mounted via NFS./pool/sao_rtdc and /home
GPFSGPFS storage mounted via InfiniBand fabric./scratch/sao
SSDLocal solid state disks.* NA
* NOTE: Users must transfer their data to and from the SSDs in their job script.

See full details at Storage Systems


Test Details

To compare the speeds of these disks we ran 10 repeats of a parellel job with 35 nodes and 500GB of requested memory. This job took a calibrated CASA measurement set and ran the partition and tclean commands to produce a final image.

You can view the job scripts for SSD and GPFS/NFS.
The python script calling CASA can be found here.


Results

TypeAverage timeMedian timeSpread
NFS52 mins47 mins31 - 73 mins.
GPFS33 mins27 mins18 - 60 mins.
SSD16 mins15 mins11 - 23 mins.


While SSDs are clearly the fastest option, there are significant time gains to be had from the GPFS disks. The majority of jobs run on the GPFS were on the shorter side, with the longer run times being less common. This is in contrast to the NFS disks where the opposite was true.

CENTER FOR ASTROPHYSICS | HARVARD & SMITHSONIAN
60 GARDEN STREET, CAMBRIDGE, MA 02138