[GPU][TESTS] Fix race condition in GpuCacheDirWithDotsParamTest

[zhanmyz]

Details:

• Fix race condition in GpuCacheDirWithDotsParamTest that causes intermittent [GPU] Failed to write 4 bytes to stream! Wrote 0 in CI parallel execution.

Description of the issue (symptom, root-cause, how it was resolved)

• Symptom: CacheDirDotVariants/GpuCacheDirWithDotsParamTest.smoke_PopulateAndReuseCache/1 fails intermittently in CI with:

  Check 'written_size == size' failed at binary_buffer.hpp:27:
  [GPU] Failed to write 4 bytes to stream! Wrote 0
  

• Root Cause:

  • The test constructs cacheDir from std::hash<std::string>{}(test_name) + GetParam(), which is fully deterministic — every process resolves to the same on-disk path (7815f5b1e52eaecf/test_encoder/test_encoder.encrypted/).
  • When gtest-parallel dispatches multiple workers concurrently, one process's removeDir(cacheDir) in SetUp()/TearDown() deletes the .blob file while another process is mid-write via write_cache_entry() → export_model() → sputn().
  • On Docker overlay2 (CI), the unlink invalidates the open fd's writable backing, causing sputn() to return 0.

• Resolution:

  • Replace the deterministic hash prefix with ov::test::utils::generateTestFilePrefix(), which incorporates thread ID and timestamp to guarantee a unique cache directory per process invocation.

The code and line that caused this issue (if it is not changed directly)

openvino/src/plugins/intel_gpu/tests/functional/subgraph_tests/serialize.cpp

Lines 75 to 94 in ea2a3c9

	void SetUp() override {
	std::stringstream ss;
	ss << std::hex << std::hash<std::string>{}(std::string(::testing::UnitTest::GetInstance()->current_test_info()->name()));
	// Base (no trailing slash first)
	cacheDir = ss.str() + GetParam();
	// Clean previous
	ov::test::utils::removeFilesWithExt(cacheDir, "blob");
	ov::test::utils::removeFilesWithExt(cacheDir, "cl_cache");
	ov::test::utils::removeDir(cacheDir);
	core.set_property(ov::cache_dir(cacheDir));
	}
	void TearDown() override {
	ov::test::utils::removeFilesWithExt(cacheDir, "blob");
	ov::test::utils::removeFilesWithExt(cacheDir, "cl_cache");
	ov::test::utils::removeDir(cacheDir);
	}

Reproduction step and snapshot (if applicable)

# Single process — always passes:
ov_gpu_func_tests --gtest_filter="CacheDirDotVariants/GpuCacheDirWithDotsParamTest.smoke_PopulateAndReuseCache/1"

# 30 parallel processes — ~30% failure rate:
for i in $(seq 1 30); do
  (ov_gpu_func_tests --gtest_filter="...smoke_PopulateAndReuseCache/1" --gtest_repeat=3 > run_$i.log 2>&1) &
done; wait
grep -l FAILED run_*.log | wc -l  # → 9/30 failed

Checklist

• [v] Is it a proper fix? (not a workaround)
• Did you include test case for this fix, if necessary?
  • No new test needed — the existing test now correctly uses a unique path and will no longer race.
• [v] Did you review existing test that can be extended to cover this scenario? The fix is within the existing test itself.

Tickets:

• CVS-182616

[Copilot]

Pull request overview

Fixes an intermittent CI failure in the Intel GPU functional test suite by eliminating a cross-process cache directory collision in GpuCacheDirWithDotsParamTest during parallel execution.

Changes:

• Replace a deterministic cache directory name (hash of test name) with ov::test::utils::generateTestFilePrefix() to ensure per-invocation uniqueness.
• Prevent concurrent gtest-parallel workers from deleting each other’s cache directories/files during SetUp()/TearDown() cleanup.