[llvm-dev] RFC: Sanitizer-based Heap Profiler

[Wenlei He via llvm-dev]

This is a big undertaking with good potential but also some uncertainty on how effective such optimizations are for larger workloads, so really appreciate the pioneering effort in LLVM.

We (facebook) are very interested in this too. I've reached out to David and Teresa a while ago about this, and was going to wait for the RFC before having more detailed discussions. But now that we're discussing it, here’s my two cents about the responsibility division between compiler and allocator, and the API.

I think it'd be beneficial if we let compiler do more heavy lighting instead of relying heavily on allocator. If we rely on less magic inside an allocator, we will likely benefit more users who may use different allocators. Otherwise there's a risk that the compiler part may be too coupled with a specific allocator, which limits the overall effectiveness of PGHO outside of that allocator.

This also affects what we want to expose in the new API for hinting the allocator (e.g. provide grouping or arena-like hint computed by compiler vs. passing a number of factors through the API which would help compute that inside allocator). With a general, stable API, hope we won't need to change API when we want to take more runtime info (temporal etc., even just for experiments) into account, or when we improve and leverage more from compiler analysis (I agree that in the long run we should improve compiler analysis).

I've talked with jemalloc folks on our side, and we're flexible to API changes. In this case, it makes sense to avoid abstraction overhead from wrappers.

Looking forward to the RFC and more discussions on this.

Thanks,
Wenlei

From: llvm-dev <llvm-dev-bounces at lists.llvm.org> on behalf of Xinliang David Li via llvm-dev <llvm-dev at lists.llvm.org>
Date: Thursday, July 8, 2021 at 9:55 AM
To: Andrey Bokhanko <andreybokhanko at gmail.com>
Cc: llvm-dev <llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] RFC: Sanitizer-based Heap Profiler


On Thu, Jul 8, 2021 at 8:03 AM Andrey Bokhanko <andreybokhanko at gmail.com<mailto:andreybokhanko at gmail.com>> wrote:
Hi Teresa,

One more thing, if you don't mind.

On Tue, Jul 6, 2021 at 12:54 AM Teresa Johnson <tejohnson at google.com<mailto:tejohnson at google.com>> wrote:
We initially plan to use the profile information to provide guidance to the dynamic allocation runtime on data allocation and placement. We'll send more details on that when it is fleshed out too.

I played with the current implementation, and became a bit concerned if the current data profile is sufficient for an efficient data allocation optimization.


First, there is no information on temporal locality -- only total_lifetime of an allocation block is recorded, not start / end times -- let alone timestamps of actual memory accesses. I wonder what criteria would be used by data profile-based allocation runtime to allocate two blocks from the same memory chunk?

First, I think per-allocation start-end time should be added to approximate temporal locality.

Detailed temporal locality information is not tracked is by design for a various of reasons:

1.  This can be done with static analysis. The idea is for the compiler to instrument a potentially hot access region and profile the start and end address of the accessed memory regions. This information can be combined with the regular heap profile data. In profile-use phase, the compiler can perform access pattern analysis and produce affinity graph

2.  We try to make use of existing allocator runtime (tcmalloc) for locality optimization. The runtime has been tuned for years to have the most efficient code for fast-path allocation.  For hot allocation sites, adding too much overhead (e.g. via wrapper etc) can lead to overhead that totally eat up the gains from the locality optimization;

3. tcmalloc currently uses size class based partitioning, which makes co-allocation of small objects of different size classes impossible. Even for objects with the same type/size, due to the use of free lists, there is no guarantee that consecutively allocated objects are placed together.

4. a bump-pointer allocator has its own sets of problems -- when not used carefully, it can lead to huge memory waste due to fragmentation.  In reality it only helps grouping for initial set of allocations when pointer bumps continuously -- during stable state, the allocations will also be all over the place and no contiguity can be guaranteed.

This is why initially we focus more coarse grain locality optimization -- 1) co-placement to improve DTLB performance and 2) improving dcache utilization using only lifetime and hotness information.

Longer term, we need to beef up compiler based analysis -- objects with the exact life times can be safely co-allocated via compiler based transformation. Also objects with similar lifetimes can be co-allocated without introducing too much fragmentation.


Thanks,

David


Second, according to the data from [Savage'20], memory accesses affinity (= space distance between temporarily close memory accesses from two different allocated blocks) is crucial: figure #12 demonstrates that this is vital for omnetpp benchmark from SPEC CPU 2017.

Said this, my concerns are based essentially on a single paper that employs specific algorithms to guide memory allocation and measures their impact on a specific set of benchmarks. I wonder if you have preliminary data that validates sufficiency of the implemented data profile for efficient optimization of heap memory allocations?

References:
[Savage'20] Savage, J., & Jones, T. M. (2020). HALO: Post-Link Heap-Layout Optimisation. CGO 2020: Proceedings of the 18th ACM/IEEE International Symposium on Code Generation and Optimization, https://doi.org/10.1145/3368826.3377914<https://doi.org/10.1145/3368826.3377914>

Yours,
Andrey

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210713/95b6d907/attachment.html>