[RFC] Implementing the BHive methodology in llvm-exegesis

Hi all,

In a recent IISWC paper, we’ve proposed BHive - a new methodology for benchmarking arbitrary basic blocks that has several advantages over the one currently used in llvm-exegesis. In particular, the new methodology:

  • automatically handles memory accesses in the basic block, without the need to manually annotate live-ins,
  • maps all memory addresses accessed by the basic block to the same page, significantly reducing the probability of cache misses during benchmarking,
  • the benchmarked code runs in a separate process, reducing risks of compromising the monitor process memory,
  • computes the throughput in a way that subtracts away the effects of the scaffolding code.

A possible challenge is increased complexity of the code: BHive uses a separate process to run the benchmarked basic block and changes memory mapping of the process to ensure that all memory accesses lead to the same page. Most operating systems have the necessary APIs, but these may differ significantly. In particular, the Windows API for memory mapping and process creation/control is very different from the Unix world. Initially, we might be able to support the new methodology only on Linux and Unix-like systems.

Before we start the implementation, we would like to collect feedback on the proposed design:

  • We’re planning to implement the methodology as a new implementation of BenchmarkRunner::FunctionExecutor that will exist alongside the current runner. The existing functionality will be preserved, and the user will be able to select the benchmark runner using a command-line flag.
  • We’re considering using the LLDB API to control the execution of the benchmarking process in a platform-independent way.

You can find a more detailed proposal here. A stand-alone Linux implementation of the methodology used in the paper is available on GitHub.

Comments and suggestions are most welcome!

Ondrej and Tom

Hi Ondrej, Tom,

This is very exciting. We’re not doing a very good job on memory instructions right now, it would be really cool to be able to measure them better. Comments inline.

In a recent IISWC paper, we’ve proposed BHive - a new methodology for benchmarking arbitrary basic blocks that has several advantages over the one currently used in llvm-exegesis. In particular, the new methodology:

  • automatically handles memory accesses in the basic block, without the need to manually annotate live-ins,
  • maps all memory addresses accessed by the basic block to the same page, significantly reducing the probability of cache misses during benchmarking,
  • the benchmarked code runs in a separate process, reducing risks of compromising the monitor process memory,
  • computes the throughput in a way that subtracts away the effects of the scaffolding code.

I’ve never actually seen a case where the scaffolding code had much influence on the results (at least on X86), especially in loop mode. However, I can see some value in snippet mode (not generated code mode): this allows the snippet code to exhaust all available registers and still be measurable.

A possible challenge is increased complexity of the code: BHive uses a separate process to run the benchmarked basic block and changes memory mapping of the process to ensure that all memory accesses lead to the same page. Most operating systems have the necessary APIs, but these may differ significantly. In particular, the Windows API for memory mapping and process creation/control is very different from the Unix world. Initially, we might be able to support the new methodology only on Linux and Unix-like systems.

Though I think it’s fine to have linux only as an initial implementation, I think there should be a clear plan to support windows: there are people in the LLVM community who are using llvm-exegesis on windows (e.g. folks at Sony). Note that you might be able to reuse some code in LLVM: compiler-rt already has an abstraction layer in “WindowsMMap.c” on top of MapViewOfFile.

Before we start the implementation, we would like to collect feedback on the proposed design:

  • We’re planning to implement the methodology as a new implementation of BenchmarkRunner::FunctionExecutor that will exist alongside the current runner. The existing functionality will be preserved, and the user will be able to select the benchmark runner using a command-line flag.

LGTM.

  • We’re considering using the LLDB API to control the execution of the benchmarking process in a platform-independent way.

I think it’s a great idea to avoid introducing any other external dependencies.

Hi Clement,

thanks for the feedback!

In a recent IISWC paper, we’ve proposed BHive - a new methodology for benchmarking arbitrary basic blocks that has several advantages over the one currently used in llvm-exegesis. In particular, the new methodology:

  • automatically handles memory accesses in the basic block, without the need to manually annotate live-ins,
  • maps all memory addresses accessed by the basic block to the same page, significantly reducing the probability of cache misses during benchmarking,
  • the benchmarked code runs in a separate process, reducing risks of compromising the monitor process memory,
  • computes the throughput in a way that subtracts away the effects of the scaffolding code.

I’ve never actually seen a case where the scaffolding code had much influence on the results (at least on X86), especially in loop mode. However, I can see some value in snippet mode (not generated code mode): this allows the snippet code to exhaust all available registers and still be measurable.

Yes, our main goal is benchmarking arbitrary basic blocks, where we do not control the register allocation.

A possible challenge is increased complexity of the code: BHive uses a separate process to run the benchmarked basic block and changes memory mapping of the process to ensure that all memory accesses lead to the same page. Most operating systems have the necessary APIs, but these may differ significantly. In particular, the Windows API for memory mapping and process creation/control is very different from the Unix world. Initially, we might be able to support the new methodology only on Linux and Unix-like systems.

Though I think it’s fine to have linux only as an initial implementation, I think there should be a clear plan to support windows: there are people in the LLVM community who are using llvm-exegesis on windows (e.g. folks at Sony). Note that you might be able to reuse some code in LLVM: compiler-rt already has an abstraction layer in “WindowsMMap.c” on top of MapViewOfFile.

Thanks for the pointers! That said, replacing mmap is relatively straightforward. The difficult part is replacing munmap, which does not have a direct equivalent on Windows and you need to query the system for all mapped blocks, and then unmap them one by one. This is a very specific functionality, and I’d be surprised if someone implemented that.

Hi all,

I’m sorry to revive such an old thread. Due to lack of time, we did not make progress as fast as we planned. We started building a prototype based on the original proposal, but we found a couple of blockers:

  • we found it very difficult to implement the interaction between the llvm-exegesis process and the child process running the benchmarked code using the LLDB API.
  • in the meantime, the MIT team continued development of the BHive algorithm, and replaced most of the assembly with C. The new code is simpler and easier to port to other architectures.

Based on our experience, we’re considering a simpler approach compared to the original proposal:

  • if possible, we will use the same C-oriented design as the latest version of the tool developed at MIT.
  • we will focus on a Linux and x86-64 implementation first, with porting to other architectures (but not operating systems) in mind. This would allow us to depend on the stable and well-defined Linux syscall interface. To our best knowledge, llvm-exegesis is already limited to Linux because of its dependence on the Linux perf subsystem, so this does not create any new portability restrictions.
  • we would use ptrace as a simpler and more powerful alternative to LLDB. It is a syscall, so it does not introduce any new external library dependencies.
  • in the same spirit, we will depend on the mmap and munmap syscalls rather than on their abstractions.

Let us know what you think!

Best regards

Ondrej

Hi all,

This is an exciting development.

BHive code base now supports both x86-64 and ARM timing. It would be great if these developments are integrated into LLVM-exegesis in a more portable manner.

From the perspective of the compiler research community, this would enable us to use precise timing data inside LLVM IR passes if we want more precision than what TTI supports. We will not have to worry about segmentation faults when timing memory heavy basic blocks if BHive’s timing methodology is adopted. Further, this will also pave way to interesting new directions in developing new performance models that are tightly coupled with the LLVM infrastructure.

Therefore, I welcome and support this contribution by Ondrej and his colleagues in getting the BHive timing infrastructure embedded within the LLVM eco-system.

Best,
Charith.