Batch download jobs

tools/interop_matrix/run_interop_matrix_tests.py

@@ -203,23 +203,17 @@ def _generate_test_case_jobspecs(lang, runtime, release, suite_name):
     return job_spec_list
 
 
-# TODO: Return a spec and then parallelize.
 def _pull_image_for_lang(lang, image, release):
     """Pull an image for a given language form the image registry."""
     cmdline = [
         'time gcloud docker -- pull %s && time docker run --rm=true %s /bin/true'
         % (image, image)
     ]
-    spec =  jobset.JobSpec(cmdline=cmdline,
+    return jobset.JobSpec(cmdline=cmdline,
                            shortname='pull_image_{}'.format(image),
                            timeout_seconds=_PULL_IMAGE_TIMEOUT_SECONDS,
                            shell=True,
-                           # TODO: Pull out to constant.
                            flake_retries=2)
-    num_failures, resultset = jobset.run([spec],
-                                         newline_on_success=True,
-                                         maxjobs=1)
-    return not num_failures
 
 
 def _test_release(lang, runtime, release, image, xml_report_tree, skip_tests):
@@ -263,17 +257,35 @@ def _run_tests_for_lang(lang, runtime, images, xml_report_tree):
     skip_tests = False
     total_num_failures = 0
 
-    # TODO: Do more intelligent chunking.
-    for release, image in images:
-        if not skip_tests and not _pull_image_for_lang(lang, image, release):
+    max_pull_jobs = min(args.jobs, _MAX_PARALLEL_DOWNLOADS)
+    max_chunk_size = max_pull_jobs
+    chunk_count = (len(images) + max_chunk_size) // max_chunk_size
+
+    for chunk_index in range(chunk_count):
+        chunk_start = chunk_index * max_chunk_size
+        chunk_size = min(max_chunk_size, len(images) - chunk_start)
+        chunk_end = chunk_start + chunk_size
+        pull_specs = []
+        if not skip_tests:
+            for release, image in images[chunk_start:chunk_end]:
+                pull_specs.append(_pull_image_for_lang(lang, image, release))
+
+        # NOTE(rbellevi): We batch docker pull operations to maximize
+        # parallelism, without letting the disk usage grow unbounded.
+        pull_failures, _ = jobset.run(pull_specs,
+                                      newline_on_success=True,
+                                      maxjobs=max_pull_jobs)
+        if pull_failures:
             jobset.message(
                 'FAILED',
                 'Image download failed. Skipping tests for language "%s"' % lang,
                 do_newline=True)
             skip_tests = True
-        total_num_failures += _test_release(lang, runtime, release, image, xml_report_tree, skip_tests)
+        for release, image in images[chunk_start:chunk_end]:
+            total_num_failures += _test_release(lang, runtime, release, image, xml_report_tree, skip_tests)
         if not args.keep:
-            _cleanup_docker_image(image)
+            for _, image in images[chunk_start:chunk_end]:
+                _cleanup_docker_image(image)
     if not total_num_failures:
         jobset.message('SUCCESS', 'All {} tests passed'.format(lang), do_newline=True)
     else: