Benchmark data for gRPC + xDS. vs envoy

[Gaurav Poothia]

Hello,

I saw a talk by Mark Roth from envoycon that talked about gRPC proxyless mesh having superior QPS per cpu second and latency than envoy all of which is of course expected.

Can anyone pls share results/setup from benchmarks around these two metrics? 

It would be great to understand perf benefits more deeply.

Thanks!

Gaurav

[Srini Polavarapu]

Hi,

The gRPC team did a one-time perf benchmarking to get some general idea. A comprehensive and continuous benchmarking plan is on the roadmap. In the ad hoc test, we tested gRPC 1.30 C++ xDS stack against Enovy that was compiled with -c opt and -fno-omit-frame-pointer from the 1.14.1 tag. Envoy was run with logging turned off entirely and with a default concurrency setting, which creates one thread per CPU. This resulted in messages being balanced across 8 threads in our set up. We were interested in the cost of a query in terms of CPU-seconds, i.e., how much CPU time is required on the client side (i.e. client process + sidecar) to transmit a single request. Load was varied from 1K to 22K QPS with 1K-byte payload.

Since this was not a comprehensive test and real world mileage depends on many things, we don't want to publish data from this test but in general you can expect to see 1.5-3x CPU savings in networking cost, i.e., the more network intensive your application is, the higher the benefits. We didn't test latency or memory utilization but you can find latency data in Istio benchmarking. 

[Nemanja Marjanovic]

Hi Srini,

I was wondering if I could maybe ask you a few questions regarding  grpc proxyless service mesh?

Thanks,

Nemanja 

[sanjay...@google.com]

Hi Namanja

Feel free to post your questions here and someone from the team will try to answer them (since Srini is no longer in the gRPC team)