Scalability with number of IAA engines

[liam0215]

I ran some benchmarks dispatching jobs to the IAA with a single work queue, 8 cores, and sweeping the engines assigned to the work queue from 1 to 8. With one engine I was able to achieve an inbound throughput on the IAA (checked using pcm_accel) of ~1.9GB/s, with 2 engines 3.8GB/s, and then it stayed flat at 3.8GB/s as I added more engines. I tested with input sizes up to 512MB and using up to 32 threads. The bottleneck was encountered with a thread/input size/core count much lower than what I described, but these are the maximum I used for the benchmark. Is this expected? My impression from some recent papers on the on-chip Intel accelerators is that the maximum throughput should be closer to 30GB/s.

[voyvoz]

It looks like you have configured the devices but have not assigned the engines to groups. You can verify this by running accel-config list and checking for entries like "ungrouped engines" or similar.
Check the example configurations from QPL (configs) for your setup.

[liam0215]

It looks like you have configured the devices but have not assigned the engines to groups. You can verify this by running accel-config list and checking for entries like "ungrouped engines" or similar. Check the example configurations from QPL (configs) for your setup.

Thanks for replying voyvoz. All the engines have been correctly assigned to a group for the shared workqueue. As evidence of being done correctly, there is a jump in throughput when going from 1 assigned engine to 2 as expected. However, beyond 2 there is no difference. What is the expected maximum inbound throughput for compression for a single shared workqueue with 8 engines assigned? For any type of source data is fine, I would just like to see a number that I can try to replicate.