CI: Benchmark x86_64-based EC2 instances using PMU instead of PERF

Previous benchmarks on x86_64-based EC2 instances indicated some
accuracy in the performance measurements using the `perf` kernel
module. Performance measurements using `rdtsc`, in turn, seem to
work more reliably.

This commit changes the benchmarking CI for x86_64 instances
to use `rdtsc` to obtain the cycle counter, and no longer `perf`.

To allow this the choice between PMU and PERF first needs to be
made an additional parameter to the EC2 benchmarking workflow.

Signed-off-by: Hanno Becker <[email protected]>

.github/workflows/bench.yml

@@ -72,36 +72,43 @@ jobs:
             ec2_ami: ubuntu-latest (aarch64)
             archflags: -mcpu=cortex-a76 -march=armv8.2-a
             cflags: -DFORCE_AARCH64
+            perf: PERF
           - name: Graviton3
             ec2_instance_type: c7g.medium
             ec2_ami: ubuntu-latest (aarch64)
             archflags: -march=armv8.4-a+sha3
             cflags: -DFORCE_AARCH64
+            perf: PERF
           - name: Graviton4
             ec2_instance_type: c8g.medium
             ec2_ami: ubuntu-latest (aarch64)
             archflags: -march=armv9-a+sha3
             cflags: -DFORCE_AARCH64
+            perf: PERF
           - name: AMD EPYC 4th gen (c7a)
             ec2_instance_type: c7a.medium
             ec2_ami: ubuntu-latest (x86_64)
             archflags: -mavx2 -mbmi2 -mpopcnt -maes
             cflags: -DFORCE_X86_64
+            perf: PMU
           - name: Intel Xeon 4th gen (c7i)
             ec2_instance_type: c7i.large
             ec2_ami: ubuntu-latest (x86_64)
             archflags: -mavx2 -mbmi2 -mpopcnt -maes
             cflags: -DFORCE_X86_64
+            perf: PMU
           - name: AMD EPYC 3rd gen (c6a)
             ec2_instance_type: c7a.medium
             ec2_ami: ubuntu-latest (x86_64)
             archflags: -mavx2 -mbmi2 -mpopcnt -maes
             cflags: -DFORCE_X86_64
+            perf: PMU
           - name: Intel Xeon 3rd gen (c6i)
             ec2_instance_type: c7i.large
             ec2_ami: ubuntu-latest (x86_64)
             archflags: -mavx2 -mbmi2 -mpopcnt -maes
             cflags: -DFORCE_X86_64
+            perf: PMU
     uses: ./.github/workflows/bench_ec2_reusable.yml
     if: github.repository_owner == 'pq-code-package' && (github.event.label.name == 'benchmark' || github.ref == 'refs/heads/main')
     with:
@@ -112,4 +119,5 @@ jobs:
       opt: ${{ matrix.opt.value }}
       store_results: ${{ github.repository_owner == 'pq-code-package' && github.ref == 'refs/heads/main' }} # Only store optimized results
       name: "${{ matrix.target.name }}${{ (!matrix.opt.value && ' (no-opt)') || ''}}"
+      perf: ${{ matrix.target.perf }}
     secrets: inherit

.github/workflows/bench_ec2_reusable.yml

@@ -34,6 +34,10 @@ on:
         type: boolean
         description: Runs with optimized code if enabled.
         default: true
+      perf:
+        type: string
+        description: Method by which clock cycles should be measured (PMU | PERF)
+        default: PERF
       store_results:
         type: boolean
         description: Indicates if results should be pushed to github pages
@@ -121,7 +125,7 @@ jobs:
           cflags: ${{ inputs.cflags }}
           archflags: ${{ inputs.archflags }}
           opt: ${{ inputs.opt }}
-          perf: PERF
+          perf: ${{ inputs.perf }}
           store_results: ${{ inputs.store_results }}
           bench_extra_args: ${{ inputs.bench_extra_args }}
           gh_token: ${{ secrets.AWS_GITHUB_TOKEN }}
@@ -148,7 +152,7 @@ jobs:
           cflags: ${{ inputs.cflags }}
           archflags: ${{ inputs.archflags }}
           opt: ${{ inputs.opt }}
-          perf: PERF
+          perf: ${{ inputs.perf }}
           store_results: ${{ inputs.store_results }}
           bench_extra_args: ${{ inputs.bench_extra_args }}
           gh_token: ${{ secrets.AWS_GITHUB_TOKEN }}