[testutils] Optimize the runtime of the method verify_no_packet_any for large port sets #231
Conversation
AntonHryshchuk
force-pushed
the
update_verify_no_packet_any
branch
from
51712d8 to
6735621
Compare
| return True # PASS - packet not observed within timeout window | ||
|
|
||
| for port in ports: | ||
| result = dp_poll( |
There was a problem hiding this comment.
you still validate it port-by-port, and not in parallel unless I don't understand the code
There was a problem hiding this comment.
Correct, the validation is still per port, but the key change is that the timeout is now global, not per-port.
Each port is polled in a non-blocking way into a loop bounded by a single timeout window.
It's not parallel, but it reduces the runtime from O(N × timeout) to O(timeout) and still detects packets on any port. Parallel way is not really possible here because the poll operates on a shared dataplane. Polling it from multiple threads not providing additional benefits and can lead to race conditions.
… scale ports number Signed-off-by: AntonHryshchuk <antonh@nvidia.com>
AntonHryshchuk
force-pushed
the
update_verify_no_packet_any
branch
from
6735621 to
3d402ef
Compare
| if port in ports: | ||
| print("verifying packet on port device", device_number, "port", port) | ||
| verify_no_packet(test, pkt, (device, port), timeout=timeout) | ||
| ports = list(ports) |
There was a problem hiding this comment.
We shouldn't change the original loop logic:
for device, port in ptf_ports():
if port in ports:
It's possible that a port in ports is not in ptf_ports(). Looks like there will be an exception in dp_poll if the port is not in ptf_ports().
|
|
||
| # Small sleep to avoid busy-looping and excessive CPU usage. | ||
| # Also gives dataplane threads time to enqueue incoming packets. | ||
| time.sleep(0.01) |
There was a problem hiding this comment.
Can we have a bigger value? 0.01s looks still too short.
|
|
||
|
|
||
| def verify_no_packet_any(test, pkt, ports=[], device_number=0, timeout=None): | ||
| def verify_no_packet_any(test, pkt, ports=[], device_number=0, timeout=1): |
There was a problem hiding this comment.
The default value of timeout was ptf.ptfutils.default_negative_timeout, we should keep it the same.
| while True: | ||
| remaining = timeout - (time.monotonic() - start) | ||
| if remaining <= 0: | ||
| return True # PASS - packet not observed within timeout window |
There was a problem hiding this comment.
Only "return" is enough here, which is the same as before.
3 participants
Optimize verify_no_packet_any method for large port sets
The original implementation waits for the full timeout on each port
sequentially, resulting in O(N * timeout) runtime (e.g. 500ports → 50 seconds, with 0.1 second of default timeout).
This change introduces a single global timeout and repeatedly polls all ports
using non-blocking dp_poll(timeout=0), ensuring:
Also adds a small sleep to avoid busy looping and reduce CPU usage.