fix(ci): always run on cncf x86_64 node.

Since we need to guarantee same-machine reproducibility for the perf reports,
it is better to run perf CI on our own node.

Also, do not upload perf data for PR since to be actually useful, we should
`perf archive` it and upload the archive. But then it would become too much large.

Signed-off-by: Federico Di Pierro <[email protected]>

.github/workflows/perf.yml

@@ -13,7 +13,7 @@ concurrency:
 
 jobs:
   perf-libs-linux-amd64:
-    runs-on: 'ubuntu-22.04'
+    runs-on: [ "self-hosted", "linux", "X64" ]
     steps:
       - name: Checkout Libs ⤵️
         uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # v4.1.1
@@ -32,20 +32,12 @@ jobs:
           make unit-test-libsinsp -j4
 
       - name: Run Perf
+        shell: bash
         run: |
           cd build
-          sudo perf record --call-graph dwarf ./libsinsp/test/unit-test-libsinsp
+          sudo perf record --call-graph dwarf -o perf.data -q libsinsp/test/unit-test-libsinsp
 
-      - name: Archive perf report (PR)
-        if: github.event_name == 'pull_request'
-        uses: actions/upload-artifact@a8a3f3ad30e3422c9c7b888a15615d19a852ae32 # v3.1.3
-        with:
-          name: perf_report
-          retention-days: 1 # just 1 day because they are big and there can be multiple ones (ie: on each push to the PR)
-          path: build/perf.data
-          if-no-files-found: error
-
-      - name: Archive perf report (master)
+      - name: Archive master perf report
         if: github.event_name == 'push'
         uses: actions/upload-artifact@a8a3f3ad30e3422c9c7b888a15615d19a852ae32 # v3.1.3
         with: