[BEAM-8944] Python SDK harness performance degradation with UnboundedThreadPoolExecutor

• Type: Bug
• Status: Resolved
• Priority: P1
• Resolution: Fixed
• Affects Version/s: 2.18.0
• Fix Version/s: 2.22.0
• Component/s: sdk-py-harness
• Assignee: Luke Cwik
• Reporter: Yichi Zhang
• Created: 11/Dec/19 02:23
• Updated: 29/May/20 22:02
• Resolved: 07/May/20 15:32

Description

We are seeing a performance degradation for python streaming word count load tests.

 

After some investigation, it appears to be caused by swapping the original ThreadPoolExecutor to UnboundedThreadPoolExecutor in sdk worker. Suspicion is that python performance is worse with more threads on cpu-bounded tasks.

 

A simple test for comparing the multiple thread pool executor performance:

 

def test_performance(self):
   def run_perf(executor):
     total_number = 1000000
     q = queue.Queue()

    def task(number):
       hash(number)
       q.put(number + 200)
       return number

    t = time.time()
     count = 0
     for i in range(200):
       q.put(i)

    while count < total_number:
       executor.submit(task, q.get(block=True))
       count += 1
     print('%s uses %s' % (executor, time.time() - t))

   with UnboundedThreadPoolExecutor() as executor:
     run_perf(executor)
   with futures.ThreadPoolExecutor(max_workers=1) as executor:
     run_perf(executor)
   with futures.ThreadPoolExecutor(max_workers=12) as executor:
     run_perf(executor)

Results:

<apache_beam.utils.thread_pool_executor.UnboundedThreadPoolExecutor object at 0x7fab400dbe50> uses 268.160675049
<concurrent.futures.thread.ThreadPoolExecutor object at 0x7fab40096290> uses 79.904583931
<concurrent.futures.thread.ThreadPoolExecutor object at 0x7fab400dbe50> uses 191.179054976
```

Profiling:
UnboundedThreadPoolExecutor:

1 Thread ThreadPoolExecutor:

12 Threads ThreadPoolExecutor:

Issue Links

• is caused by: BEAM-8151 Allow the Python SDK to use many many threads
• is related to: BEAM-10158 [Python] Reuse a shared unbounded thread pool
• links to: GitHub Pull Request #10387
• links to: GitHub Pull Request #10430
• links to: GitHub Pull Request #11590