Benchmarking Ray's object storage for bandwidth and latency

I’m trying to reproduce the numbers for bandwidth and latency shown at the top of page 10 of the paper available here:

Does anyone here maybe have the code and setup instructions required to see these numbers?
Any kind of pointer is much appreciated since I’d like to avoid reinventing the wheel.

cc @pcmoritz can you help with this?

when i’m running a simple python application that writes and reads to/from the built-in object storage, i somehow can’t get more than ~30 MB/sec. So what i’m actually looking for is a way to demonstrate that much more bandwidth possible – i’m fine if that is not the original source ccde used for the paper up there, but I’d be interested in some way to reach these numbers.

Is there someone who could help there?

From our microbenchmark, we have around 15GB/s for the throughput. The script is here ray/ at master · ray-project/ray · GitHub.

Also, you should make sure you are using /dev/shm properly for writing your objects

Microbenchmark result; ray/microbenchmark.txt at master · ray-project/ray · GitHub look single client put gigabytes per second 15.51 +- 12.57