[llvm-dev] [RFC] Context-sensitive Sample PGO with Pseudo-Instrumentation

[Wenlei He via llvm-dev]

Thanks for the thoughtful questions, David. See my answers inline.

Thanks,
Wenlei

From: Xinliang David Li <davidxl at google.com>
Date: Friday, August 7, 2020 at 1:24 PM
To: Wenlei He <wenlei at fb.com>
Cc: "llvm-dev at lists.llvm.org" <llvm-dev at lists.llvm.org>, Wei Mi <wmi at google.com>, Hongtao Yu <hoy at fb.com>
Subject: Re: [RFC] Context-sensitive Sample PGO with Pseudo-Instrumentation

Wenlei, Thanks for the interesting proposal! please see my replies inline below.

On Fri, Aug 7, 2020 at 11:28 AM Wenlei He <wenlei at fb.com<mailto:wenlei at fb.com>> wrote:
Hi All,
Our team at Facebook is building a new context-sensitive Sample PGO as an alternative to the existing AutoFDO. We’d like to share our motivation, propose a new design, and reveal preliminary results on benchmarks. We will refer to the proposed design as CSSPGO in this RFC.

The new CSSPGO leverages simultaneous LBR and stack sampling to construct a full context-sensitive profile.


Can you share more details on this? LBR only has 32 entries, so it won't give you full call context, so stack unwinding is needed. What is the overhead you see in production environment?

[wenlei] We are not worried about overhead in production environment as the sampling rate there is extremely low. We did measure locally however, for stack sampling and level 2 PEBS on top of regular LBR sampling, the overheads isn’t very noticeable still, but I don’t have numbers at hand.

It doesn’t rely on previous inlining like today’s AutoFDO to get context-sensitive profile, and it also doesn’t need a separate post-inline context-sensitive profile like CSPGO.

What is the sample profile data size impact with the full context information?

[wenlei] Text CS profile is normally around 1x-10x of regular profile size, with all live context included. We plan to trim cold context, which we expect to bring the size down in a meaningful way. Another source of size increase is the context string, which could contain duplicated mangle names (can be very long for C++ templated code), but should be very compressible with the built-in compression support from extended binary profile. We will move to extended binary format, and leverage the compression support if needed. We can also consider more efficient fixed-length integer context representation (similar to rolling hash).


In addition, we introduced pseudo-instrumentation for more accurate mapping from binary samples back to IR, similar to instrumentation PGO, but without any measure-able runtime overhead that is usually associated with instrumentation.


Is CSSPGO inherently dependent upon pseudo-probe or is it orthogonal? I hope that it is the latter :)

[wenlei] They’re orthogonal. Context-sensitive SPGO can work without pseudo-probe and still use dwarf. Our plan is to keep context-sensitive SPGO working w/ and w/o pseudo-probe functionality-wise, but we only look at perf and tune with the two combined.


We have a functioning implementation for the new CSSPGO now. Initial results on SPEC2006 shows ~2% geomean performance win on top of AutoFDO (with MonoLTO and NewPM) and ~4% .text size reduction at the same time.


Motivation
AutoFDO is a big success as it lowers the entry barrier for PGO significantly while still delivering substantial performance boost. However, there’s still a gap between AutoFDO and its instrumentation counterpart. From several failed internal attempts to improve AutoFDO, we realized that the bottleneck of AutoFDO lies in its profile quality. With the current level of profile quality, it’s difficult to reap more performance win because good heuristics are often limited by inferior profile. That prompted a systemic effort to investigate and improve AutoFDO framework. Specifically, we’re trying to handle the two biggest sources of profile quality issues:


  1.  AutoFDO relies on a limited context-sensitive profile collected based on previous inlining. Therefore it can only replay or prune the previous inlining. With the main CGSCC inliner, post-inline counts are not accurate due to scaling of context-less profile, which affects the effectiveness of later passes such as profile-guided code layout.

Acknowledge of the limitation here.


  1.
  2.  Dwarf line and discriminator info aren’t always well-maintained throughout the compilation, thus using them as anchors to map binary samples back to the IR can sometimes be inaccurate, which leads to inferior profile quality and limits PGO performance.

I think we need more quantification of the impact of using debug information for matching purposes:  How much performance are left on the table due to this, and are they fixable issues or not.

[wenlei] The first table in the result section is comparing pseudo-probe with AutoFDO and Instr. PGO, all with inlining turned off. So that’s a quantitative assessment of the effectiveness of pseudo-probe. It’s hard to assess performance benefit though, because PGO performance is a function of profile quality and heuristic. Currently heuristics are tuned to cope with the profile quality we have, so it may not do justice for profile quality improvements that pseudo-probe brings us.

One example is how FDO inliner evaluates call site. It checks callee’s total sample count instead of callee’s entry count. This is less than ideal, but we couldn’t fix it due to profile quality issues – we can’t reliably get inlinee’s entry count with dwarf based approach, see discussion in https://reviews.llvm.org/D60086. That problem is solved with pseudo-probe, but until we change the inliner, we won’t see perf win from that particular profile quality improvement. There are other similar cases too, and that’s why we used profile quality metric instead of performance to assess pseudo-probe.

Some of the issues may be fixable with dwarf info maintenance, but the engineering cost to find and fix all issues are non-trivial. We think maintaining anchor as IR is a more sustainable alternative than maintaining anchor as metadata.



  1.
To lift the above limitations, we’d like to propose an alternative design that consists of two components: 1) Context-sensitive sample PGO, 2) Sample to IR mapping using pseudo probes. The goal is to further improve sample PGO performance while maintaining usability and keeping training runtime overhead at zero. In addition, we hope the CSSPGO framework can also open up opportunities for new optimizations with more stringent requirements on profile quality.



CSSPGO is a very attractive optimization by itself.  Can you provide more motivation for the pseudo probes?

[wenlei] One way to look at the combination of pseudo-probe and context-sensitive sample PGO is that, the former brings sample PGO closer to instrumentation PGO, and the latter to sample PGO is like the two-stage CSPGO, or even post-link optimizer to instrumentation PGO. These are two orthogonal problems that needs separate solutions.

There’re other secondary motivations for pseudo-probe as well beyond its profile quality benefits that I didn’t mention earlier:
1). Stale profile detection. With line numbers, it’s hard to detect and react to stale profile. Pseudo-probes are tied to blocks so it’s effectively using CFG as carrier for profile, which makes stale profile detection easier.
2). Resilience to source changes. We’ve seen cases where deleting a single line of comment caused a 8% perf regression for a large service because it completely messed up profile annotation for a critical path. That will not happen with pseudo-probe – any source change not altering CFG will be tolerated without perf impact.
3). Possibility of offline count inference. We have an experiment that encodes edges alongside with probes (blocks), so more sophisticated offline count inference algorithm is possible to further improve profile quality. Our algorithm researchers are working on new profile inference solution now.



Context-sensitive Sample PGO
The effectiveness of BOLT, Propeller and CSPGO all demonstrated the importance of context-sensitive profile for PGO. However there are two limitations with the existing approaches.

  1.  The current solutions focus on leveraging a context-sensitive profile to attain an accurate post-inline profile that helps achieve a better code layout, but do not use the context-sensitive profile to drive better inlining.
  2.  The current solutions involve multiple training processes and profiles (e.g. a post-inline profile for CSPGO, or a post-link profile for BOLT and Propeller), which incurs higher operational cost and complicates the build and release workflow.
We propose a full context-sensitive sample profiling infrastructure that utilizes both LBR and call stack samples at the same time to synthesize a profile with a full context sensitivity. The key advantage is that rather than relying on previous inlining or a separate profile, the profile collected with the new approach will have full calling contexts recovered from both inlined and not inlined call sites. To achieve an accurate post-inline profile, a separate profile is no longer needed. Instead, the post-inline profile can be directly derived from adjusting the input profile based on all inline decisions. The richer context-sensitive profile also enables better inline decisions. The infrastructure has two key components listed below.

Synthesizing context-sensitive LBR with a virtual unwinder
To make sample PGO’s input profile context aware, we need to know the call stack of each LBR fall through path. That is done by sampling LBR and call stack simultaneously. With that, each sample will contain a call stack in addition to LBR entries. We use level 2 PEBS to control sampling skid so that the leaf frame from stack sample aligns with leaf frame from LBR. The raw call stack sample describes the calling context for the leaf LBR entry. In addition, by unwinding “call” and “return” (including implicit ones from inlinee) from LBR entries backwards on top of raw stack samples, we can recover the calling context for each of the LBR entries from the sample, thus synthesizing context-sensitive LBR profile.

We can then generate context-sensitive sample PGO profile using the context-sensitive LBR profile. In the new profile, a function’s profile becomes a collection of profiles, each representing a profile for a given calling context.


Sounds good -- see the overhead question posted at the beginning.



Context-sensitive FDO/PGO framework in LLVM
In order to leverage context-sensitive profile for inlining, and to maintain accurate post-inline counts, we introduced SampleContextTracker which is a layer sitting in between input profile and the profile used to annotate CFG for optimizations. We also introduced the notion of base profile which is the merged profile for function’s profiles from any outstanding (not inlined) context, and context profile which is a function's profile for a given calling context. The framework includes four simple APIs for updating and query profiles:

Query API:

  *   getBaseSamplesFor: Query base profile by function name.
  *   getContextSamplesFor: Query context profile by calling context and function name.
Update API:

  *   MarkContextSamplesInlined: When a function is inlined for a given calling context, we need to mark the context profile for that context as inlined. This is to make sure we don't include inlined context profile when synthesizing the base profile.
  *   PromoteMergeContextSamplesTree: When a function is not inlined for a given calling context, we need to promote the context profile tree to be top-level context. This preserves the child context under that function so later inline decisions for calls originating from that not inlined function will still be driven by an accurate context profile.
These APIs are used by SampleProfileLoader’s inlining, for better inline decisions and better post-inline counts. For optimal results, the new infrastructure needs to work with a top-down FDO inliner. We added top-down FDO inlining under a switch, and the switch is turned on by default in upstream recently. There’re a few other improvements for the FDO inliner that we plan to upstream soon.

The profile data should be usable by the SCC inliner as well. In the bottom up inlining, as the function gets inline further up in the call chain, the inline instance has few incoming contexts to merge.

[wenlei] Yes, we intentionally introduced the SampleContextTracker abstraction that is decoupled from SampleProfileLoader, so it can work with both FDO inliner and SCC inliner. But we expect FDO inliner to take over more inlining for CSSPGO because the FDO inliner is no longer a replay inliner now. And it’s good as top-down inline helps with specialization which is important for context-sensitive inlining.



Pseudo-instrumentation for sample to IR mapping
Being able to profile production binaries is a key advantage of AutoFDO over Instrumentation PGO, but it also comes with a big challenge. While using line number and discriminator as anchor for profile mapping incurs zero run time overhead for AutoFDO, it’s not as accurate as instrumented probes. This is because the instrumented probes are part of the IR, rather than metadata attached to the IR like !dbg. That has two implications: 1) it’s easier to maintain IR than metadata for optimization passes; 2) probe blocks some CFG transformations that can mess up profile correlation.

With the proposed pseudo instrumentation, we can achieve most of the benefit of instrumentation PGO in little runtime overhead. We instrument each basic block with a pseudo probe associated with the block Id. Unlike in PGO instrumentation where a counter is implemented as a persisting operation such as atomic read/write or runtime helper call, a pseudo probe is implemented as a dedicated intrinsic call with IntrInaccessibleMemOnly flag. The intrinsic comes with most of the semantics of a PGO counter but is much less optimization-intrusive.

The pseudo probe intrinsic calls are on the IR throughout the optimization and code generation pipeline and are materialized as a piece of binary data stored in a separate .pseudo_probe data section.

How are these information maintained? Blocks can be eliminated or cloned in many optimization passes: jump threading, taildup, unrolling, peeling etc.  For instance, how to handle the block that is merged into another? Does it lose samples because of this?

[wenlei] They are just maintained as part of IR, like any other instructions, without special care. The key difference is they’re part of IR instead of metadata attached to IR. We can categorize relevant CFG transformations into 1) duplication, 2) merge and removal.
For any duplication, tail/head dup, unrolling, probe will be duplicated along with other instructions, and we don’t need duplication factor that was used by dwarf-based approach, because counts from duplicated probes will be added together naturally. For merge and removal, IntrInaccessibleMemOnly flag will block it, similar to real probes.

Pseudo-probe is a framework that is tunable. Depending on the semantic we put on the intrinsic, it can be as heavy as real probe, or as light as a label. IntrInaccessibleMemOnly is a carefully chosen semantic based on our experiments that balances run time overhead and profile quality – it doesn’t incur measure-able overhead even though it still blocks merging and removal, we didn’t see measure-able overhead from SPEC or a large internal workload. But the profile quality improvement is measure-able as the 1st table in result section shows.



The section is then used to map binary samples back to blocks of CFG during profile generation. There are also no real machine instructions generated for a pseudo probe and the.pseudo_probe section won’t be loaded into memory at runtime, therefore they should incur very little runtime overhead. As a fact, we see no measure-able performance impact from pseudo-instrumentation itself on SPEC2006 or big internal workload.

How large are the probe sections?

[wenlei] About 10% of binary size, another 2% if we encode CFG edges in addition to probes/blocks.


Pseudo-instrumentation integration and Pass Ordering
One implication from pseudo-probe instrumentation is that the profile is now sensitive to CFG changes. We now defect stale profiles for sample PGO via CFG checksum, instead of just using it. However, the potential downside is that CFG may change between different versions of the compiler even if the source code is unchanged. To solve that problem, we perform the pseudo instrumentation very early in the pre-LTO pipeline, before any CFG transformation. This ensures that the CFG instrumented and annotated is stable. We added SampleProfileProber that performs the pseudo instrumentation and runs independent of profile annotation.

A new switch -fpseudo-probe-for-profiling is added to enable sample PGO with pseudo instrumentation, similar to -fdebug-info-for-profiling for AutoFDO. Input profile is still provided through the same switch used by today’s AutoFDO, namely -fprofile-sample-use, and the profile loader will automatically decide how to load and annotate profile depending on whether input profile is dwarf-based or pseudo-probe based.


Can you compare the source change tolerance of pseudo probe based approach vs debug info based approach?

[wenlei] Pseudo-probe should be more resilient to source changes. See my reply for motivation of pseudo-probe. Pseudo-probe will be able to tolerate source changes as long as they don’t alter CFG. On the contrary, changes that delete a comment and shift line offset can cause perf churn with line-based approach. We've been bitten by this a few times – people making comment only change during holiday freeze only to find surprising perf regression due to AutoFDO 😊. It also opens up possibility of fuzzy CFG matching when there’s a CFG mutation due to source change to make it even more resilient.



New profile format and profile generation
We extend current profile format in order to be able to represent a full context-sensitive profile and also encode pseudo-probe info. This is done without drastically diverging from today’s AutoFDO profile format so that existing tools and libraries can be reused with minor changes (e.g. llvm-profdata, profiler reader and writer).

For a context-sensitive profile, we extend the profile format by changing the function profile header line to include calling context in addition to function name. With today’s AutoFDO, we have a single profile header for each function to represent its accumulative profile. E.g. This is the profile header for foo, with 1290 total samples, and 74 header samples.

foo:1290:74

For CSSPGO, we will have multiple profile headers for a single function’s profile, each representing profile for a specific calling context as shown below. CSSPGO profile header is bracketed to differentiate from today’s AutoFDO.

[main:12 @ bar:3 @ foo]:279:54
[main:19 @ zoo:7 @ foo]:1011:20


sounds good.



With calling context encoded in the function header, we no longer need a nested function profile for inlinees. Instead, a context profile will be represented uniformly using context strings in the function profile header, regardless of whether the calls in the context are inlined or not. The flat structure makes sure that context profile is easily indexable. The change is mostly transparent to tools like llvm-profdata. Support for binary profile format has not been added yet, but should be easy to do.



It is still useful to annotate (as least with comment line) that a profile is for top level function or inline instance.

[wenlei] Agreed, and that’s in our plan too - we need that for tuning purpose.


For pseudo-probe, we repurposed the line to count map of AutoFDO profile to be block Id to count map. This only changes the interpretation of profile content rather than the representation, hence all reader/writer helpers can be reused.

There's a new profile generation tool, llvm-profgen, with the virtual winder implemented for context-sensitive profiling, and uses the .pseudo_probe section to map binary profile to pre-opt CFG profile. Since profile generation is a critical piece of the workflow, we’d like to propose to include the tool as part of LLVM, alongside with llvm-profdata.


Preliminary Results
To quantitatively assess profile quality improvement brought by pseudo-instrumentation, we introduce a profile quality metric. We measure the metric by first annotating an optimized binary with the MIR block execution counts derived from a profile. The binary is then sampled and the counts are compared against the annotation. The weighted relative delta is used as an indicator for profile quality (lower is better).

Table below shows the profile quality metric for SPEC2006. We can see from the numbers that the profile quality of pseudo-instrumentation sample PGO is much better than AutoFDO and close to instrumentation PGO.

Profile quality metric
Baseline AutoFDO
Instrumentation PGO
Sample PGO w/ Pseudo Instrumentation
SPEC2006
24.58%
15.70%
16.21%


Instrumentation PGO does not have context sensitivity, so I would expect it scores worse than CSSPGO. Do you know why it is better here?

[wenlei] This is for evaluating effectiveness of pseudo-probe exclusively. We have all inlining turned off for this experiment, and this is without context-sensitive profile for Sample PGO either, so the comparison should be fair, and Instrumentation PGO should be the upper bound.



We also measured performance and code size on SPEC2006 with CSSPGO. The measurement was done with MonoLTO and new pass manager, with tuning for FDO inliner to accommodate context-sensitive profile, and used training dataset for both pass1 and pass2. The result shows ~2% performance win on top of today’s AutoFDO, with ~4% .text reduction, see table below.

SPEC2006
Performance
Code Size
AutoFDO over LTO
CSSPGO
Over LTO
CSSPGO over AutoFDO
AutoFDO over LTO
CSSPGO
Over LTO
CSSPGO over AutoFDO
Geomean Delta %
6.80%
8.70%
2.04%
11.17%
6.66%
4.06%

While the SPEC2006 benchmark suite is different from large workloads, we think the results demonstrated the potential of CSSPGO and served its purpose for proof of concept. We plan to continue tuning and start to evaluate larger internal workloads soon, and we’d like to upstream our work. Feedbacks are welcomed!



What is the performance win with peudo-probe alone?

[wenlei] We don’t have numbers for pseudo-probe along. As I mentioned earlier, profile quality improvement may not translate directly to perf win without heuristic changes. That’s why we evaluate pseudo-probe exclusively with profile quality metric. The hope is that it will open up opportunity for better optimizations. E.g. it could potentially help the Machine Function Splitting pass too. That said, pseudo-probe does bring extra win for CSSPGO comparing to line-based CSSPGO for some benchmarks, but we didn’t tune CSSPGO with line-based profile, so we didn’t aggregate numbers as the comparison isn’t fair either.


thanks,

David



Thanks,
Wenlei & Hongtao


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200807/c214c6cd/attachment-0001.html>