I agree with the multiple modes strategy, it would be great to enable tracking both sema + index and index performance without any additional cost, if the framework would be generic enough that would be even better! The other responses are inline.
Thanks for your responses!
I realize now that I should have been more specific when it comes to completion latency. We’re currently interested in sema completion latency, but the infrastructure that I would like to set up will support latency with the completion results obtained from the index as well.
Essentially, for a completion test-case we would like to have the option to run it in two / three modes:
- just sema completion
- index completion or sema + index completion
Note that we don’t have to test a completion test-case in all modes, so we could just have a sema based completion test.
This way we’ll be able to identify the regressions in a particular component (sema vs index) in a better way. Do you think this idea works for you?
More responses inline:
Such test-suite might be very useful and it’d be great to have it. As Eric mentioned, I am working on pulling benchmark library into LLVM. Although I fell behind over the past week due to the complications with libc++ (you can follow the thread here: http://lists.llvm.org/pipermail/llvm-dev/2018-August/125176.html).
Thanks! Do you a general idea of how you would like to use the benchmarking library?
I’ve looked into benchmark usage in libc++ and test-suite, but they weren’t very helpful because they seem to be very specific there. I’ve started looking into pulling the library into LLVM (https://reviews.llvm.org/D50894) but I have few concerns there.
I’m mainly interested in a more complete test that we could run using some sort of harness and whose results can be fed into LNT.
Can you please elaborate on what you mean by feeding results into LNT? Are you thinking about controlling the latency and failing the “benchmark tests” as soon as the latency is beyond some limit or are you interested in building LNT targets which you can run along unittests?
Eric, Ilya and I have been discussing a possible “cheap” solution - a tool which user can feed a compilation database and which could process some queries (maybe in YAML format, too). This would allow a realistic benchmark (since you could simply feed LLVM codebase or anything else with the size you’re aiming for) and be relatively easy to implement. The downside of such approach would be that it would require some setup effort. As an alternative, it might be worth feeding YAML symbol index instead of the compilation commands, because currently the global symbol builder is not very efficient. I am looking into that issue, too; we have few ideas what the performance bottlenecks in global-symbol-builder can be and how to fix them, hopefully I will make the tool way faster soon.
Note that sema latency is something we also need to take into consideration, as it’s always part of code completion flow, with or without index.
In the long term, however, I think the LLVM Community is also interested in benchmarking other tools which exist under the LLVM umbrella, so I think that opting in for the Benchmark approach would be more beneficial. Having an infrastructure based on LNT that we could run either on some buildbots or locally would be even better. The downside is that it might turn out to be really hard to maintain a realistic test-suite, e.g. storing YAML dump of the static index somewhere would be hard because we wouldn’t want 300+ Mb files in the tree but hosting it somewhere else and downloading would also potentially introduce additional complexity. On the other hand, generating a realistic index programmatically might also be hard.
I don’t have a strong opinion for how the index should be stored. However, I think it’s helpful to breakdown this problem into different categories, and look at three kinds of indexing data sets:
- index data set that’s derived from a part of the LLVM umbrella (llvm/clang/test-suite/whatever).
=> One possible solution: this index can be rebuilt on every run.
Yes, but that unfortunately takes too long at the moment. I started looking into that and fixed a YAML serialization performance problem (https://reviews.llvm.org/D50839), but there are few other bottlenecks left.
- index data set that’s derived form a project outside of the LLVM umbrella.
=> One possible solution: This index can be stored as an archive of YAML files in one of the LLVM repos.
That’s one of doing it, right, but I’m not sure any LLVM repo would like to store a 300 Mb YAML file and update it over time. However, I don’t know if there are already any cases like this and whether it might be acceptable.
- auto generated index data?
While this might be the most appealing option, the generation of a realistic index might turn out to be hard. However, I think we should have couple of artificial indices in the benchmarks, it might be beneficial.
It would probably be valuable to have different kinds of index data sets.
Agreed, that would also mean more coverage.
Relatively “cheap” solutions which I’m thinking about are:
- Recording user session and mirroring the input file to Clangd to measure the performance. That would eliminate the complexity of creating a realistic benchmark without investing too much effort into the benchmark itself. However, it might turn out to be hard to track the performance contributions of individual components. Also, I’m not sure if it’s generic enough.
- Creating a tool which would accept YAML symbol dump, build an index and get a set of requests (e.g. from another file) to measure total completion latency. That solves most of my problems, but is tied to the index testing usecase which is not enough for comprehensive performance tracking.
I unfortunately didn’t get any good idea of how to build comprehensive performance tracking pipeline yet, e.g. how to continuously get index for (e.g.) LLVM, adjust buildbots, measure performance, ensure that it’s realistic, etc.
Having said that, convenient infrastructure for benchmarking which would align with the LNT and wouldn’t require additional effort from the users would be amazing and we are certainly interested in collaboration. What models of the benchmarks have you considered and what do you think about the options described above?
For the sema based completion latency tracking I would like to start off with two simple things to get some basic infrastructure working:
- C++ test-case: measuring sema code-completion latency (with preamble) in a file from a fixed revision of Clang.
- ObjC test-case: similar to above, some ObjC code with a portion.
One issue is that it the system headers that will be used are not static, which leads to issues like the baseline might be out of date when the SDK on the GreenDragon bots is updated.
Ideally I would use some harness based on compile commands. Each test file would have a compilation command entry in the database.
I was also thinking that the test command could be fed into Clangd using LSP itself. Similarly to how code-completion is requested in Clangd’s regression test, we could write a test that would send in the LSP commands into Clangd. Or maybe the test harness could generate them from some sort of test description (e.g. test completion at these locations at that file).
The latency could be measured by scanning the output of the run of Clangd with CLANGD_TRACE.
The test harness would then capture the result and upload it to LNT. A subsequent bot would check for big regressions (e.g. +10%) against the baseline (or previous result).
Sounds good to me!