[dpdk-dev] [PATCH v5 07/10] autotest: properly parallelize unit tests

[Reshma Pattan <reshma.pattan@intel.com>]

Now that everything else is in place, we can run unit tests in a
different fashion to what they were running as before. Previously,
we had all autotests as part of groups (largely obtained through
trial and error) to ensure parallel execution while still limiting
amounts of memory used by those tests.

This is no longer necessary, and as of previous commit, all tests
are now in the same group (still broken into two categories). They
still run one-by-one though. Fix this by initializing child
processes in multiprocessing Pool initialization, and putting all
tests on the queue, so that tests are executed by the first idle
worker. Tests are also affinitized to different NUMA nodes using
taskset in a round-robin fashion, to prevent over-exhausting
memory on any given NUMA node.

Non-parallel tests are executed in similar fashion, but on a
separate queue which will have only one pool worker, ensuring
non-parallel execution.

Support for FreeBSD is also added to ensure that on FreeBSD, all
tests are run sequentially even for the parallel section.

Cc: stable@dpdk.org

Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
---
 test/test/autotest.py        |   6 +-
 test/test/autotest_runner.py | 277 +++++++++++++++++++++++++++----------------
 2 files changed, 183 insertions(+), 100 deletions(-)

diff --git a/test/test/autotest.py b/test/test/autotest.py
index ae27daef7..12997fdf0 100644
--- a/test/test/autotest.py
+++ b/test/test/autotest.py
@@ -36,8 +36,12 @@ def usage():
 
 print(cmdline)
 
+# how many workers to run tests with. FreeBSD doesn't support multiple primary
+# processes, so make it 1, otherwise make it 4. ignored for non-parallel tests
+n_processes = 1 if "bsdapp" in target else 4
+
 runner = autotest_runner.AutotestRunner(cmdline, target, test_blacklist,
-                                        test_whitelist)
+                                        test_whitelist, n_processes)
 
 runner.parallel_tests = autotest_data.parallel_test_list[:]
 runner.non_parallel_tests = autotest_data.non_parallel_test_list[:]
diff --git a/test/test/autotest_runner.py b/test/test/autotest_runner.py
index d6ae57e76..36941a40a 100644
--- a/test/test/autotest_runner.py
+++ b/test/test/autotest_runner.py
@@ -6,16 +6,16 @@
 from __future__ import print_function
 import StringIO
 import csv
-import multiprocessing
+from multiprocessing import Pool, Queue
 import pexpect
 import re
 import subprocess
 import sys
 import time
+import glob
+import os
 
 # wait for prompt
-
-
 def wait_prompt(child):
     try:
         child.sendline()
@@ -28,22 +28,47 @@ def wait_prompt(child):
     else:
         return False
 
-# run a test group
-# each result tuple in results list consists of:
-#   result value (0 or -1)
-#   result string
-#   test name
-#   total test run time (double)
-#   raw test log
-#   test report (if not available, should be None)
-#
-# this function needs to be outside AutotestRunner class
-# because otherwise Pool won't work (or rather it will require
-# quite a bit of effort to make it work).
+
+# get all valid NUMA nodes
+def get_numa_nodes():
+    return [
+        int(
+            re.match(r"node(\d+)", os.path.basename(node))
+            .group(1)
+        )
+        for node in glob.glob("/sys/devices/system/node/node*")
+    ]
+
+
+# find first (or any, really) CPU on a particular node, will be used to spread
+# processes around NUMA nodes to avoid exhausting memory on particular node
+def first_cpu_on_node(node_nr):
+    cpu_path = glob.glob("/sys/devices/system/node/node%d/cpu*" % node_nr)[0]
+    cpu_name = os.path.basename(cpu_path)
+    m = re.match(r"cpu(\d+)", cpu_name)
+    return int(m.group(1))
+
+
+pool_child = None  # per-process child
 
 
-def run_test_group(cmdline, prefix, target, test):
+# we initialize each worker with a queue because we need per-pool unique
+# command-line arguments, but we cannot do different arguments in an initializer
+# because the API doesn't allow per-worker initializer arguments. so, instead,
+# we will initialize with a shared queue, and dequeue command-line arguments
+# from this queue
+def pool_init(queue, result_queue):
+    global pool_child
+
+    cmdline, prefix = queue.get()
     start_time = time.time()
+    name = ("Start %s" % prefix) if prefix != "" else "Start"
+
+    # use default prefix if no prefix was specified
+    prefix_cmdline = "--file-prefix=%s" % prefix if prefix != "" else ""
+
+    # append prefix to cmdline
+    cmdline = "%s %s" % (cmdline, prefix_cmdline)
 
     # prepare logging of init
     startuplog = StringIO.StringIO()
@@ -54,24 +79,60 @@ def run_test_group(cmdline, prefix, target, test):
         print("\n%s %s\n" % ("=" * 20, prefix), file=startuplog)
         print("\ncmdline=%s" % cmdline, file=startuplog)
 
-        child = pexpect.spawn(cmdline, logfile=startuplog)
+        pool_child = pexpect.spawn(cmdline, logfile=startuplog)
 
         # wait for target to boot
-        if not wait_prompt(child):
-            child.close()
+        if not wait_prompt(pool_child):
+            pool_child.close()
+
+            result = tuple((-1,
+                            "Fail [No prompt]",
+                            name,
+                            time.time() - start_time,
+                            startuplog.getvalue(),
+                            None))
+            pool_child = None
+        else:
+            result = tuple((0,
+                            "Success",
+                            name,
+                            time.time() - start_time,
+                            startuplog.getvalue(),
+                            None))
+    except:
+        result = tuple((-1,
+                        "Fail [Can't run]",
+                        name,
+                        time.time() - start_time,
+                        startuplog.getvalue(),
+                        None))
+        pool_child = None
 
-            return -1, "Fail [No prompt]", "Start %s" % prefix,\
-                   time.time() - start_time, startuplog.getvalue(), None
+    result_queue.put(result)
 
-    except:
-        return -1, "Fail [Can't run]", "Start %s" % prefix,\
-               time.time() - start_time, startuplog.getvalue(), None
+
+# run a test
+# each result tuple in results list consists of:
+#   result value (0 or -1)
+#   result string
+#   test name
+#   total test run time (double)
+#   raw test log
+#   test report (if not available, should be None)
+#
+# this function needs to be outside AutotestRunner class because otherwise Pool
+# won't work (or rather it will require quite a bit of effort to make it work).
+def run_test(target, test):
+    global pool_child
+
+    if pool_child is None:
+        return -1, "Fail [No test process]", test["Name"], 0, "", None
 
     # create log buffer for each test
     # in multiprocessing environment, the logging would be
     # interleaved and will create a mess, hence the buffering
     logfile = StringIO.StringIO()
-    child.logfile = logfile
+    pool_child.logfile = logfile
 
     # make a note when the test started
     start_time = time.time()
@@ -81,7 +142,7 @@ def run_test_group(cmdline, prefix, target, test):
         print("\n%s %s\n" % ("-" * 20, test["Name"]), file=logfile)
 
         # run test function associated with the test
-        result = test["Func"](child, test["Command"])
+        result = test["Func"](pool_child, test["Command"])
 
         # make a note when the test was finished
         end_time = time.time()
@@ -109,15 +170,6 @@ def run_test_group(cmdline, prefix, target, test):
         result = (-1, "Fail [Crash]", test["Name"],
                   end_time - start_time, logfile.getvalue(), None)
 
-    # regardless of whether test has crashed, try quitting it
-    try:
-        child.sendline("quit")
-        child.close()
-    # if the test crashed, just do nothing instead
-    except:
-        # nop
-        pass
-
     # return test results
     return result
 
@@ -137,7 +189,7 @@ class AutotestRunner:
     blacklist = []
     whitelist = []
 
-    def __init__(self, cmdline, target, blacklist, whitelist):
+    def __init__(self, cmdline, target, blacklist, whitelist, n_processes):
         self.cmdline = cmdline
         self.target = target
         self.binary = cmdline.split()[0]
@@ -146,6 +198,8 @@ def __init__(self, cmdline, target, blacklist, whitelist):
         self.skipped = []
         self.parallel_tests = []
         self.non_parallel_tests = []
+        self.n_processes = n_processes
+        self.active_processes = 0
 
         # log file filename
         logfile = "%s.log" % target
@@ -159,11 +213,8 @@ def __init__(self, cmdline, target, blacklist, whitelist):
         self.csvwriter.writerow(["test_name", "test_result", "result_str"])
 
     # set up cmdline string
-    def __get_cmdline(self):
-        cmdline = self.cmdline
-
-        # affinitize startup so that tests don't fail on i686
-        cmdline = "taskset 1 " + cmdline
+    def __get_cmdline(self, cpu_nr):
+        cmdline = ("taskset -c %i " % cpu_nr) + self.cmdline
 
         return cmdline
 
@@ -241,6 +292,51 @@ def __filter_test(self, test):
 
         return True
 
+    def __run_test_group(self, test_group, worker_cmdlines):
+        group_queue = Queue()
+        init_result_queue = Queue()
+        for proc, cmdline in enumerate(worker_cmdlines):
+            prefix = "test%i" % proc if len(worker_cmdlines) > 1 else ""
+            group_queue.put(tuple((cmdline, prefix)))
+
+        # create a pool of worker threads
+        # we will initialize child in the initializer, and we don't need to
+        # close the child because when the pool worker gets destroyed, child
+        # closes the process
+        pool = Pool(processes=len(worker_cmdlines),
+                    initializer=pool_init,
+                    initargs=(group_queue, init_result_queue))
+
+        results = []
+
+        # process all initialization results
+        for _ in range(len(worker_cmdlines)):
+            self.__process_result(init_result_queue.get())
+
+        # run all tests asynchronously
+        for test in test_group:
+            result = pool.apply_async(run_test, (self.target, test))
+            results.append(result)
+
+        # tell the pool to stop all processes once done
+        pool.close()
+
+        # iterate while we have group execution results to get
+        while len(results) > 0:
+            # iterate over a copy to be able to safely delete results
+            # this iterates over a list of group results
+            for async_result in results[:]:
+                # if the thread hasn't finished yet, continue
+                if not async_result.ready():
+                    continue
+
+                res = async_result.get()
+
+                self.__process_result(res)
+
+                # remove result from results list once we're done with it
+                results.remove(async_result)
+
     # iterate over test groups and run tests associated with them
     def run_all_tests(self):
         # filter groups
@@ -253,77 +349,60 @@ def run_all_tests(self):
                    self.non_parallel_tests)
         )
 
-        # create a pool of worker threads
-        pool = multiprocessing.Pool(processes=1)
-
-        results = []
+        parallel_cmdlines = []
+        # FreeBSD doesn't have NUMA support
+        numa_nodes = get_numa_nodes()
+        if len(numa_nodes) > 0:
+            for proc in range(self.n_processes):
+                # spread cpu affinity between NUMA nodes to have less chance of
+                # running out of memory while running multiple test apps in
+                # parallel. to do that, alternate between NUMA nodes in a round
+                # robin fashion, and pick an arbitrary CPU from that node to
+                # taskset our execution to
+                numa_node = numa_nodes[self.active_processes % len(numa_nodes)]
+                cpu_nr = first_cpu_on_node(numa_node)
+                parallel_cmdlines += [self.__get_cmdline(cpu_nr)]
+                # increase number of active processes so that the next cmdline
+                # gets a different NUMA node
+                self.active_processes += 1
+        else:
+            parallel_cmdlines = [self.cmdline] * self.n_processes
 
-        # whatever happens, try to save as much logs as possible
-        try:
+        print("Running tests with %d workers" % self.n_processes)
 
-            # create table header
-            print("")
-            print("Test name".ljust(30) + "Test result".ljust(29) +
-                  "Test".center(9) + "Total".center(9))
-            print("=" * 80)
+        # create table header
+        print("")
+        print("Test name".ljust(30) + "Test result".ljust(29) +
+              "Test".center(9) + "Total".center(9))
+        print("=" * 80)
 
-            # print out skipped autotests if there were any
-            if len(self.skipped):
-                print("Skipped autotests:")
+        if len(self.skipped):
+            print("Skipped autotests:")
 
-                # print out any skipped tests
-                for result in self.skipped:
-                    # unpack result tuple
-                    test_result, result_str, test_name, _, _, _ = result
-                    self.csvwriter.writerow([test_name, test_result,
-                                             result_str])
+            # print out any skipped tests
+            for result in self.skipped:
+                # unpack result tuple
+                test_result, result_str, test_name, _, _, _ = result
+                self.csvwriter.writerow([test_name, test_result, result_str])
 
-                    t = ("%s:" % test_name).ljust(30)
-                    t += result_str.ljust(29)
-                    t += "[00m 00s]"
+                t = ("%s:" % test_name).ljust(30)
+                t += result_str.ljust(29)
+                t += "[00m 00s]"
 
-                    print(t)
+                print(t)
 
-            # make a note of tests start time
-            self.start = time.time()
+        # make a note of tests start time
+        self.start = time.time()
 
+        # whatever happens, try to save as much logs as possible
+        try:
             if len(self.parallel_tests) > 0:
                 print("Parallel autotests:")
-                # assign worker threads to run test groups
-                for test_group in self.parallel_tests:
-                    result = pool.apply_async(run_test_group,
-                                              [self.__get_cmdline(),
-                                               "",
-                                               self.target,
-                                               test_group])
-                    results.append(result)
-
-            # iterate while we have group execution results to get
-            while len(results) > 0:
-
-                # iterate over a copy to be able to safely delete results
-                # this iterates over a list of group results
-                for group_result in results[:]:
-
-                    # if the thread hasn't finished yet, continue
-                    if not group_result.ready():
-                        continue
-
-                    res = group_result.get()
-
-                    self.__process_result(res)
-
-                    # remove result from results list once we're done with it
-                    results.remove(group_result)
+                self.__run_test_group(self.parallel_tests, parallel_cmdlines)
 
             if len(self.non_parallel_tests) > 0:
                 print("Non-parallel autotests:")
-                # run non_parallel tests. they are run one by one, synchronously
-                for test_group in self.non_parallel_tests:
-                    group_result = run_test_group(
-                        self.__get_cmdline(), "", self.target, test_group)
-
-                    self.__process_result(group_result)
+                self.__run_test_group(self.non_parallel_tests, [self.cmdline])
 
             # get total run time
             cur_time = time.time()
-- 
2.14.4