ghz VS gRPC

I once built a quick and dirty load testing tool for a public facing service we built. The tool was pretty simple - something like https://github.com/bojand/ghz but with traffic and data patterns closer to what we expected to see in the real world. We used argo-workflows to generate scale.

One thing which we noticed was that there was a considerable difference in performance characteristics based on how we parallelized the load testing tool (multiple threads, multiple processes, multiple kubernetes pods, pods forced to be distributed across nodes).

I think that when you run non-distrubuted load tests you benefit from bunch of cool things which happen with http2 and Linux (multiplexing, resource sharing etc) which might make applications seem much faster than they would be in the real world.

I'm not sure. Maybe you can write to the discussion section of the repo https://github.com/bojand/ghz/discussions

In part 1 we looked at ghz for load testing gRPC services, and now I want to cover k6, which claims to be a modern load testing tool built for developer happiness. After only a brief experience with it I can see why is that and why Grafana moved to acquire k6 earlier this year.

It should be unbound, in this particular benchmark we set ghz concurrency to 50 and connections to 5 and we don't set the rps flag of ghz (e.g. --rps=2000, from this tool)

a second container running ghz makes unary requests to the server

That's not true at all. Case in point In general, this is not a problem that AGC can solve. The language can help (something Java is admittedly particularly bad at) but even so, there'll always be avenues for leaks. That's just the nature of shared things. Interestingly, in the linked grpc case, the leaked memory is only half the problem -- AGC doesn't help at all with the leaked HTTP2 connection.

I've trialled my new image with an existing project via JLink that's heavy on Netty and gRPC the image works great (with a small tweak to exclude grpc-netty-shaded due to grpc-java#9083).

When developing an application, the vast majority of code is written by other people. We import that code and make use of it to get whatever we need done. In this case, the developer of an various android applications are using grpc-java.

`proto_library` for building the `.bin` file from protos works great. Generating stubs/messages for "all" languages does not. Each language does not want to implement gRPC rules, the gRPC team does not want to implement rules for each language. Sort of a deadlock situation. For example:

- C++: https://github.com/grpc/grpc/blob/master/bazel/cc_grpc_libra...

- Python: https://github.com/grpc/grpc/blob/master/bazel/python_rules....

- ObjC: https://github.com/grpc/grpc/blob/master/bazel/objc_grpc_lib...

- Java: https://github.com/grpc/grpc-java/blob/master/java_grpc_libr...

- Go (different semantics than all of the other): https://github.com/bazelbuild/rules_go/blob/master/proto/def...

But there's also no real cohesion within the community. The biggest effort to date has been in https://github.com/stackb/rules_proto which integrates with gazelle.

tl;dr: Low alignment results in diverging implementations that are complicated to understand for newcomers. Buff's approach is much more appealing as it's a "this is the one way to do the right thing" and having it just work by detecting `proto_library` and doing all of the linting/registry stuff automagically in CI would be fantastic.

Small clarification (to my understanding, I'm not a Java Guru) on why Java got on top - those Java implementations use something called Direct Executor. It's super performant when there's no chance of a blocking operation. But if you are to do anything more than echo service, you might be in trouble. Other implementations probably don't suffer from the same constraint. The related discussion can be found in this PR.

clone https://github.com/grpc/grpc-java

If you do streaming then the best option would be to use a so called manual flow control. You can find an example here.

Another interesting link is their official grpc-java benchmarks project, which is also used in the benchmark I've posted you.

Both JLink (gRPC#3522) and Graal have some issues; I'm especially concerned about the Serial GC in Graal so will be putting that under some stress soon to see if that confirms my suspicions. I'll also be good when some Java 16 JRE Alpine images appear as the JDK is too bloaty.