search

intel / thin-layout-optimizer

MIT License
53 stars 1 forks source link

readme

Perf Profile -> Optimized Linking Function Layout

thin-layout-optimizer is a project that takes a performance profile from a tool like perf and uses the data (particularly lbr data) to the layout of the profiled DSO(s) functions.

Currently it uses a basic hfsort algorithm (or other re-ordering algorithms) to do this. The output is a text file with some information from the profiles (including the order), which is meant to be consumed by the script finalize-order.py to create a function order list for either ld or gold.

Build

Tests

Sanitized Build

Usage

ld Setup

  1. Patches:

    • git am patches/ld/*.patch

  2. Usage(LD_ORDERING_SCRIPT):

    • export LD_ORDERING_SCRIPT=ordering.lds; ld ...

    NOTE If the .lds linker script file exists, it must be a syntactically valid linker script, otherwise ld will fail. If the fail doesn't exist it will be ignored.

  3. Usage(LD_ORDERING_SCRIPT_MAP): Map syntax:

    target:<target_name0>,<target_name1>,...<target_nameN>
DSO0 <ordering-script0>
DSO1 <ordering-script1>
...
DSON <ordering-scriptN>
    • export LD_ORDERING_SCRIPT_MAP=ordering_map; ld ...

    Note the target name is optional. If it is present, we will only using the ordering script map if the target string matches the link target for ld/gold. If it is not present, we will always use the map. The target name is the name used for the emulation (-m) option of ld (even if using gold, specify the ld emulation name). For example to make a map file for only x86_64 specify target:elf_x86_64 before any DSOs.

    Note: Prior to reaching a target: line, we will watch DSOs to any target. This also means that if target: is not present, we will match all the DSOs regardless of target.

  4. Unpatched ld: It is possible to use the finalize-order.py script to generate an ld linker script using the ld.orig target. The generated linker script will have the ordering embedded in the best guess of the default linker script. While you may have some success with this, it is not advised as some ld options may essentially end up overriden by overriding the default linker script.

gold

Without Patches

  1. Usage Without Patches:

    • gold --section-ordering-file ordering.txt ...

    NOTE The gold linker tends to do a slightly better job than ld.

With Patches

  1. Setup patches:

    -patches/ld/*.patch

  2. Usage(GOLD_ORDERING_SCRIPT) With Patches:

    • export GOLD_ORDERING_SCRIPT=ordering.gold; gold ...

  3. Usage(GOLD_ORDERING_SCRIPT_MAP) With Patches:

    • export GOLD_ORDERING_SCRIPT_MAP=ordering_map; gold ...

thin-layout-optimizer

  1. Collect a profile of the system using.

    • perf record -e cycles:u,branch-misses -j any,u -a

  2. Package the result. This will copy all the referenced DSOs + some debug file to a new tar file. thin-layout-optimizer will use the copied DSOs as its references (as opposed to the system ones which, if updated, will changed the addresses of functions and make the old data unusable).

    • python3 scripts/package.py <input:perf.data file> <dst:packaged-profile>

    <dst> will always be a .tar.gz file, even if you didn't specify .tar.gz

  3. Unpackage the results when you are ready to create the function ordering(s).

    • python3 scripts/unpackage.py <src:packaged-profile.tar.gz> <dst:unpackaged-profile dir>

    This is really just a wrapped for untarring the tar created in Step 2.

    Note: Both the package.py and unpackage.py scripts accepts a 3rd optional argumet compress. If this is provided it will run perf on the perf.data profile and compress the output (with zstd). Once you have the compressed zst files, it is fine to delete the perf.data file.

  4. Run thin-layout-optimizer.

    • thin-layout-optimizer -r <src:unpackaged-profile dir> -o <dst:dir-for-ordering-file> --save <dst:saved-state-file>

    Alternatively, if you just want to save a profile to combine with other future profiles, or just to create ordering files later you can use:

    Note: There are more options (see thin-layout-optimizer -h). For the most part just using the above command should be all you need to do.

  5. (Optional) Save/Reload From Saved States.

    • When running thin-layout-optimizer with a new perf.data profile, you can use the option --save to store the state just before call-graph creation. After creating a save-state, you can re-run thin-layout-optimizer using the --reload option to avoid the time-consuming task of processing the perf.data files. You can also combine multiple save-states with the option. For example:

    1. Saving a profile:

      • thin-layout-optimizer -r <src:unpackaged-profile dir> --save <dst:saved-state-file>

    2. Reloading save state

      • thin-layout-optimizer --reload <src:saved-state-file> -o <dst:dir-for-ordering-file>

    3. Reloading and combining multiple saved state

      • thin-layout-optimizer --reload <src:saved-state-file0>,<src:saved-state-file1>,<src:saved-state-fileN> --save <dst:combined-save-state-file>

  6. Finalize ordering for a target.

    Note: This section assumes usage of the ld/gold patches with the *_ORDERING_SCRIPT_MAP env variable.

    • Once a set of ordering files have been created the final step is to create the linker scripts to be used for re-linking. Note the ordering files can either be from a perf.data file, save states, or a combination of the two.

    • The method for creating the *_ORDERING_SCRIPT_MAP and linker scripts from the ordering files is to use the scripts/finalize-order.py script.

    • python3 scripts/finalize-order.py -i <src:dir-with-ordering-files> -t <'ld', 'ld.orig', or 'gold'> -o <dst:directory for linker scripts> -m <dst:map files for env variable> -p <perfix for paths in map file> --align-hot-n <float:0-100> --align-till <float:0-100> --alignment=<int:0-12> --align-per-dso --aliases <file:alias-file>

      • The -i argument is the directory where you saved the ordering files (the -o argument from thin-layout-optimizer)
      • The -m argument will create the file you should set to *_ORDERING_SCRIPT_MAP.
      • The -o argument will be where the actual linker scripts are stored.
      • The -p argument will set prefix for where the directory with the scripts should be searched for by the *_ORDERING_SCRIPT_MAP. I.e if you use -p ~/foobar, the *_ORDERING_SCRIPT_MAP will contain entries like <dso name> ~/foobar/<linker script for dso name>. If the -p option isn't present it will default to what was used for -o.
      • The --alignment option specifies the log2 of the alignment we want. I.e --alignment=5 (default) will align functions to 32 bytes. --alignment=10 will likewise be 1024 byte alignment.
      • The --align-hot-n option species we want to add --alignment to the top N percentage of called functions. I.e if --align-hot-n=1 and --alignment=5, we will align the functions that are in the 99% percentile of number of incoming calls to 32 bytes.
      • The --align-till option species we want to add --alignment to functions until we account of N percentage of total calls. I.e if --align-till=33 and --alignment=12, we will keep aligning functions to 4096 bytes until 33% of the total incoming edges have been accounted for.
      • The --align-per-dso option takes the above two constraints (--align-hot-n and --align-till), and instead of applying them globally, will apply them for each DSO. For example if we have two DSOs A and B with A containing 99% of the total calls. If you don't specify --align-per-dso, then we will align functions in both A/B based on their global hotness (so likely no function from B). If you do specify --align-per-dso, we will apply both alignment constraints to both A and B regardless of their relative weight.
      • The --aliases option takes a file that maps one target to another. This is useful if the commandline profiled is different than the build target, or if you want to use the same ordering file for multiple targets. Note if the --alias argument is missing, it will try to find one at the path specified by the env variable ORDERING_SCRIPT_ALIASES.

    Note: An example from running this script when targeting llvm:

    • For gold (with an explicit prefix):

      • python3 scripts/finalize-order.py -i orders/ -t gold -o llvm-orders/ -m llvm-orders-map.gold -p ~/.llvm-orders

    • For ld (without an explicit prefix):

      • python3 scripts/finalize-order.py -i orders/ -t ld -o ~/.llvm-orders/ -m llvm-orders-map.ld

    By default the script will name the actual script files with a .ld or .gold extension based on the target, so scripts for ld/gold can safely be stored to the same directory.

    NOTE: By default finalize order will emit lines to find each function at:

    1. .text.<func>
    2. .text.hot.<func>
    3. .text.cold.<func>
    4. .text.unlikely.<func>

    To skip the cold and unlikely locations, pass --no-cold to finalize-order.py.

  7. Rebuild the project.

    • For ld either pass the output order-script.lds to ld directly with the -T option, or if using the patched version set env export LD_ORDERING_SCRIPT=/path/to/order-script.lds
    • For gold you can use the builtin argument --section-ordering-file.

  8. Verify new order.

    • To verify a new order, you can use scripts/compare-order.py. This will compare the ordering in a built binary, compare it with an ordering file, and produce a "score" for how close the binary and ordering file are. The lower the score, the closer the orders are.. Note the ordering file referenced here is not the output script used in the link stage, but the assosiated file from Step 4 in directory <dir-for-ordering-file>.

      python3 scripts/compare-order.py <src:ordering file> <src:bin0> ... <src:binN>

      I.e when using for LLVM:

      python3 scripts/compare-order.py orders-new-new/ordering--home-noah-programs-opensource-llvm-dev-src-project3-build-ld-bin-clang-18.txt /home/noah/programs/opensource/llvm-dev/src/project3/build-gold-order/bin/clang-18 /home/noah/programs/opensource/llvm-dev/src/project3/build-ld-order/bin/clang-18 /home/noah/programs/opensource/llvm-dev/src/project3/build-ld/bin/clang-18
/home/noah/programs/opensource/llvm-dev/src/project3/build-gold-order/bin/clang-18 -> 12.72/12.72
/home/noah/programs/opensource/llvm-dev/src/project3/build-ld-order/bin/clang-18 -> 68.423/68.423
/home/noah/programs/opensource/llvm-dev/src/project3/build-no-order/bin/clang-18 -> 16963.985/16963.985

      Note: Generally there will be no difference between the two numbers, if there is its an indication that the profile was too minimal.

Comparing Benchmark Results

To get an idea if new linker scripts are needed, based on new
benchmark results, you can use the `compare-dir.py` script. This
script will see how different the function orders in two
directories are. Then, based on the result, you can decide if the
results warrant creating new linker scripts.

  1. Get Orders for Benchmarks:

    • thin-layout-optimizer -r <src:unpackaged-profile dir benchmark0> -o <dst:dir-for-ordering-files0> --save <dst:saved-state-file0>
    • thin-layout-optimizer -r <src:unpackaged-profile dir benchmark1> -o <dst:dir-for-ordering-files1> --save <dst:saved-state-file1>

  2. Compare Output Directories:

    • python3 scripts/compare-dir.py <dst:dir-for-ordering-files0> <dst:dir-for-ordering-files1>

    This will the difference of all common and distinct DSOs in both order directores. As well as a summary at the end.

  3. Advanced DSO Matching: By default compare-dir.py will match .so files across different versions (i.e match foo.so.1.1 with foo.so.1.2). To disable this pass --exact-version.

  4. Common DSOs Only: To only compare the common DSOs (skip the summary of DSOs that are not in both directories), use the --ignore-distinct option.

Advanced Benchmark Recombination

By default, you reload multiple save-states the save states are
all normalized. This ensure that longer running/shorter running
(but equally important) benchmarks are both weighted the same when
recombined.

  1. Normalization: Normalization is on by default. To disable to you can use either --no-normalize or --force-no-normalize. The only difference between these two options, is that --no-normalize will fail if the reloaded save states have a mixed of already normalized / not-normalized files (its pretty easy to see why this is likely not desirable). The --force-no-normalize option will stil warn in this case, but won't kill the process. An example use might be: thin-layout-optimizer --no-normalize --reload <src:saved-state-file0>,<src:saved-state-file1>,<src:saved-state-fileN> --save <dst:combined-save-state-file>

  2. Scaling: It may be desirable to scale certain benchmarks. I.e there may be a set of very important benchmarks, but also some less important ones that still provide meaningful coverage. To scale certain benchmarks (increase / decrease there importance) you can either use the --add-scale option when creating a save state, or manually edit a save state.

    • Adding a Scale By Hand: Underneath the hood, the save states a just a json file. To manually add a save state you can insert entries under the "scaling" field:

      "scaling": {
"scale": <Some Double above 0.0>
}

      Its possible the "scaling" field won't already exist in the save state. If thats the case, just add it.

      You might also see other options in the "scaling" field. For example:

      "scaling": {
"edge_normalized": true,
"func_normalized": true,
"edge_scale" : 2.5,
"func_scale" : 2.0
},

      The "edge_normalized" and "func_normalized" fields should never be modified by hand. These are for tracking which files have been normalized.

      The "edge_scale" and "func_scale" options are equivilent to "scale" but with a bit more precision as to what they apply to. To scale only weights among edges, but not function sample rate, or vice versa, you can modify/insert these fields independently. If the "scale" field exists, its value will override both "func_scale" and "edge_scale". For example with:

      "scaling": {
"edge_normalized": true,
"func_normalized": true,
"edge_scale" : 2.5,
"func_scale" : 2.0
"scale" : 3.0
},

      Both function samples and edge samples will be scaled by 3.0.

    • Adding a Scale With The Commandline: To add a scale to a single file, instead of modify by hand you can also do: thin-layout-optimizer --no-normalize --add-scale <Some Double Above 0.0> --reload <src:saved-state-file0> --save <dst:saved-state-file0-with-assosiated-scale>

      The --no-normalize is not strictly necessary, but will preserve the original values.

      Likewise, if you wish to combine multiple save states and assosiate a scale with the combined save state you can also do: thin-layout-optimizer --add-scale <Some Double Above 0.0> --reload <src:saved-state-file0>,<src:saved-state-file1>,<src:saved-state-fileN> --save <dst:saved-state-file0-N-with-assosiated-scale>

      Whether you normalize is optional again, just keep in mind once a save state has been normalized, its original precise values will have been lost.

    Using a Scale: To actually scale a file, use the --use-custom-scale commandline option. thin-layout-optimizer --no-normalize --reload --save <dst:saved-state-file0-with-assosiated-scale> --save <dst:saved-state-file0-scaled>

    The `--use-custom-scale` option will check whats in the
`"scaling"` field and use any `"scale"`, `"func_scale"`,
and/or `"edge_scale"` values to multiply the relevant metrics.

**NOTE: You <u>cannot</u> use both `--add-scale` and
`--use-custom-scale` at the same time. `--add-scale` only
assosiates a scale with the save state. It does not actually
multiply the fields.**.

Finally, you can invoke `--use-custom-scale` when recombining multiple save states:
`thin-layout-optimizer --use-custom-scale --reload <src:saved-state-file0>,<src:saved-state-file1>,<src:saved-state-fileN>  --save <dst:saved-state-file0-N-scaled>`

When combining multiple save states with `--use-custom-scale`,
the scale will be ontop of normalization. As well, each
save-state will use its own `"scaling"` field.