test(perf): split block tests across two runners

This has two reasons:
- When adding block latency tests (e.g. duplicating all existing test
  cases to also run with fio's sync workload generator), the runtime
  will exceed 1 hour, which is the buildkite pipeline timeout)
- Having the sync and async cases in the same buildkite step means that
  the A/B-testing framework will try to cross-correct between the sync
  and async engine, but comparing results from these makes no sense
  because they are completely disjoint code paths in firecracker and
  the host kernel, so there is no reason to believe that their
  regressions should be correlated.

Signed-off-by: Patrick Roy <[email protected]>

Author: roypat

.buildkite/pipeline_perf.py

@@ -16,9 +16,14 @@
 # the operating system sometimes uses it for book-keeping tasks. The memory node (-m parameter)
 # has to be the node associated with the NUMA node from which we picked CPUs.
 perf_test = {
-    "virtio-block": {
+    "virtio-block-sync": {
         "label": "💿 Virtio Block Performance",
-        "tests": "integration_tests/performance/test_block_ab.py::test_block_performance",
+        "tests": "integration_tests/performance/test_block_ab.py::test_block_performance -k 'not Async'",
+        "devtool_opts": "-c 1-10 -m 0",
+    },
+    "virtio-block-async": {
+        "label": "💿 Virtio Block Performance",
+        "tests": "integration_tests/performance/test_block_ab.py::test_block_performance -k Async",
         "devtool_opts": "-c 1-10 -m 0",
     },
     "vhost-user-block": {