SAS recently tested SAS 9.4 on GCP instances using EXAScaler Cloud by DDN: Google Marketplace - EXAScaler Cloud. The results of the testing were very favorable.
The test infrastructure consisted of 21 GCP instances defined like this:
- Three n2-standard-32 instances for the SAS clients – each with 9 TB of local SSD storage.
- Sixteen n1-standard-16 instances for the OSS hosts – each with 4.3 TB of pd-balanced (balanced persistent disks) storage
- One n1-standard-4 instance for the MGS host – with 2 TB of pd-ssd (SSD persistent disks) storage
- One n1-standard-32 instance for the MDS host – with 1 TB of pd-balanced storage.
- EXAScaler 5.1.1 was used for the testing.
For the test, the following EXAScaler tunings were applied:
- OSS tuning - Run on all OSS nodes
lctl set_param osd-ldiskfs.*.read_cache_enable=0 osd-ldiskfs.*.writethrough_cache_enable=0 obdfilter.*.brw_size=16 obdfilter.*.precreate_batch=1024
- Lustre client tuning - Run on all client nodes
lctl set_param mdc.*.max_rpcs_in_flight=128 osc.*.max_pages_per_rpc=16M osc.*.max_rpcs_in_flight=16 osc.*.max_dirty_mb=1024 llite.*.max_read_ahead_mb=2048 osc.*.checksums=0 llite.*.max_read_ahead_per_file_mb=256
And the following file system tuning was applied:
lfs setstripe -c -1 -S 4M < file system subdirectory >
rhel_iotest results:
The 1-node test gave us 221 MB per second per physical core for WRITEs and 333 MB per second per physical core for READs.
The 2-node test gave us 226 MB per second per physical core for WRITEs and 337 MB per second per physical core for READs.
The 3-node test gave us 230 MB per second per physical core for WRITEs and 325 MB per second per physical core for READs.
Please note that SAS recommends a minimum of 100 MB per second per physical core for WRITEs and READs.
As you can see, the above EXAScaler Cloud infrastructure can produce the minimum IO throughput needs of a 64-core SAS Grid application. Please reach out to DDN for more details on EXAScaler Cloud and how to configure it for optimal performance in GCP.
