Has any one developed as a preference or recommendation on SASWORK for 9.4
Looks like I have Ultra Disk and Premium SSD to choose on these machines.
For Ultra Disk, is there a guidance on determining iops? The MBPS is a bit easier but I'm not sure how we determine that with SAS
Thanks!
Azure L series instances - please make sure you use the Intel ones. There are possible issues with the AMD ones since SAS heavily uses Intel's math kernel libraries (MKL) for analytics. Sometimes the MKL libraries do not work on AMD, so performance gains from the MKL libraries will not be available on AMD systems.
As for Ultra vs Premium drives, we have not seen enough performance gains with Ultra drives to justify the additional price of Ultra drives when compared to Premium drives.
Azure L series instances - please make sure you use the Intel ones. There are possible issues with the AMD ones since SAS heavily uses Intel's math kernel libraries (MKL) for analytics. Sometimes the MKL libraries do not work on AMD, so performance gains from the MKL libraries will not be available on AMD systems.
As for Ultra vs Premium drives, we have not seen enough performance gains with Ultra drives to justify the additional price of Ultra drives when compared to Premium drives.
Looks like we should actually be using the nvme drives that are already on the box. Based on the documentation we should hit some really high speeds. What options were used by SAS when striping across the 4 nvme drives? Settings like stripsize and readahead
Tuning for SASWORK
there are SAS NOTES etc out there on this topic. I start with striping at the BUFSIZE.
@epower What OS and version are you running? If RHEL 7 or 8, I would recommend using the rhel_iotest.sh tool (https://support.sas.com/kb/59/680.html) to measure your IO throughput over the standard UNIX version. Rhel_iotest.sh is much easier to run since you only need to pass it the target file system and it handles calculating the rest of the parameters needed.
As for the nvme striping (note this also applies to RHEL), we typically recommend using lvm2 to stripe these drives together. A good place to start is with a 64 KB stripesize and a large, static value for readahead (8 MB+). I've found a 16 MB readahead is a good fit for my workload, but you can always monitor and adjust these values down the road if needed. When using lvm2, it's also important to make sure the drives ultimately get striped and not concatenated.
Let me know if you have any questions or would like some help putting together commands for setting up/striping the nvme drives.
You can find our RHEL OS tuning best practices here:
RHEL 6 & 7: https://access.redhat.com/sites/default/files/attachments/optimizingsasonrhel6and7_2.pdf
Thanks for the recommendations! We're on RHEL 8 so I will definitely be using that tool along with some benchmark programs I have written in SAS itself as well.
Happy to help! The current external version of rhel_iotest.sh only supports RHEL 6 & 7 (we're working on getting RHEL 8 support pushed). If you want to shoot me an email at Jim.Kuell@sas.com, I'd be happy to send you over a temporary version that will work with RHEL 8.
EDIT: It looks like the RHEL 8 update has been pushed to the external version after all. I'll see if I can get the doc updated to properly reflect that.
SAS provides a tool for testing IO: 51660 - Tool to test I/O throughput on UNIX platforms (sas.com)
A minimum throughput of 100mps per core is recommended.
Hi Everyone,
Just wanted to follow up (not sure what I was thinking about considering premium or ultra on the L series machine)... Anyways, we tested out the L32v3series using the NVME for work and this is what we got using the iotest script. I would say this is a huge success. Pushing 7GB/sec on the L32 right out of the box on SASWORK.
The SAS Users Group for Administrators (SUGA) is open to all SAS administrators and architects who install, update, manage or maintain a SAS deployment.
SAS technical trainer Erin Winters shows you how to explore assets, create new data discovery agents, schedule data discovery agents, and much more.
Find more tutorials on the SAS Users YouTube channel.