DPDK: 5-tuple-swap multiprocess testpmd on a single port

[mcgov]

Adds a simpler forwarding test using 2 VMs and 1 port. This test uses multiprocess DPDK to create a sender and receiver process on one VM, and a forwarder process on the other. The forwarder sends all traffic back to the first VM by swapping the mac and IP addresses.

[Copilot]

Pull request overview

This PR adds a new DPDK test that validates multiprocess testpmd functionality using a 5-tuple-swap forwarding mode with a single port setup. The test involves two VMs: one VM runs both sender and receiver processes using DPDK multiprocessing, while the other VM acts as a forwarder that swaps MAC and IP addresses to return traffic back to the sender VM.

Changes:

• Introduces multiprocess DPDK support with new enums for process roles and forwarding modes
• Adds generate_5tswap_run_info function to configure the 3-process test topology
• Updates command handling throughout to support multiple commands per node (List[str] instead of str)
• Adds new test case verify_dpdk_testpmd_5tswap_gb_hugepages_netvsc to the test suite

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 10 comments.

File	Description
lisa/microsoft/testsuites/dpdk/common.py	Adds DpdkMpRole and TestpmdForwardMode enums to support multiprocess contexts and different forwarding modes
lisa/microsoft/testsuites/dpdk/dpdktestpmd.py	Adds _generate_mp_arguments method and extends generate_testpmd_command to support multiprocess role configuration and custom core lists
lisa/microsoft/testsuites/dpdk/dpdkutil.py	Adds generate_5tswap_run_info function for 5-tuple-swap test setup, updates return types to support multiple commands per node, and modifies verify_dpdk_send_receive to handle multiple processes
lisa/microsoft/testsuites/dpdk/dpdksuite.py	Adds new test case for 5-tuple-swap multiprocess validation with 1GB hugepages

[LiliDeng]

@mcgov it failed in canonical 0001-com-ubuntu-server-jammy 22_04-lts-gen2 latest, please check it is test case issue or image issue.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 5 comments.

Comments suppressed due to low confidence (1)

lisa/microsoft/testsuites/dpdk/dpdkutil.py:662

• start_testpmd_concurrent now accepts multiple commands per DpdkTestResources, but _collect_dict_result stores results in output keyed only by the kit, so later command outputs will overwrite earlier ones for the same kit. If multi-process output is needed, change the output structure (e.g., Dict[DpdkTestResources, List[str]] or key by (kit, proc_index)), or explicitly document/rename the function to indicate only the last output per kit is retained.

    def _collect_dict_result(result: Tuple[DpdkTestResources, str]) -> None:
        output[result[0]] = result[1]

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

Comments suppressed due to low confidence (1)

lisa/microsoft/testsuites/dpdk/dpdkutil.py:662

• start_testpmd_concurrent now accepts a list of commands per test kit, but _collect_dict_result stores a single string per kit (output[result[0]] = result[1]). If more than one command is provided for the same kit, later-finishing tasks will overwrite earlier outputs non-deterministically, losing logs/results. Consider changing the output type to store a list per kit (or key by (kit, index)), and update the collector accordingly.

    command_pairs_as_tuples: List[Tuple[DpdkTestResources, str]] = []
    kits_and_commands = deque(node_cmd_pairs.items())
    for kit_and_commands in kits_and_commands:
        kit, commands = kit_and_commands
        for command in commands:
            command_pairs_as_tuples += [(kit, command)]

    def _collect_dict_result(result: Tuple[DpdkTestResources, str]) -> None:
        output[result[0]] = result[1]

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 8 comments.

Comments suppressed due to low confidence (1)

lisa/microsoft/testsuites/dpdk/dpdkutil.py:669

• start_testpmd_concurrent now supports multiple commands per test kit, but _collect_dict_result stores results in output keyed only by DpdkTestResources, so later commands overwrite earlier ones. This makes the returned output incomplete/ambiguous for multiprocess runs; consider storing a list of outputs per kit (or key by (kit, proc_id) / command index) and updating the return type accordingly.

    command_pairs_as_tuples: List[Tuple[DpdkTestResources, str]] = []
    kits_and_commands = deque(node_cmd_pairs.items())
    for kit_and_commands in kits_and_commands:
        kit, commands = kit_and_commands
        for command in commands:
            command_pairs_as_tuples += [(kit, command)]

    def _collect_dict_result(result: Tuple[DpdkTestResources, str]) -> None:
        output[result[0]] = result[1]

[Copilot]

The forwarding ratio check uses hard-coded thresholds is_close_to(0.8, 0.2) without any context for why 80% (±20%) is the expected range. Please document the rationale (e.g., expected overhead, batching behavior) or define named constants so it's clear what behavior/regression this is intended to catch.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 8 comments.

[Copilot]

run_testpmd_concurrent/start_testpmd_concurrent now expect Dict[DpdkTestResources, List[str]], but this path still stores a single string command. This will cause iteration over characters when flattening commands. Wrap this as a one-element list (and ensure any other assignments in this function follow the same List[str] convention).

[Copilot]

With multiple commands per DpdkTestResources, _collect_dict_result overwrites prior output for the same kit, losing logs/results from earlier processes. Consider changing output to Dict[DpdkTestResources, List[str]] (preferred) or appending/concatenating outputs under the same key so multi-process/multi-command runs preserve all outputs.

[mcgov]

don't do that.

[Copilot]

With multiple commands per DpdkTestResources, _collect_dict_result overwrites prior output for the same kit, losing logs/results from earlier processes. Consider changing output to Dict[DpdkTestResources, List[str]] (preferred) or appending/concatenating outputs under the same key so multi-process/multi-command runs preserve all outputs.

Suggested change

	existing_output = output.get(result[0])
	if existing_output is None:
	output[result[0]] = result[1]
	else:
	# concatenate outputs for the same kit to avoid losing earlier results
	output[result[0]] = f"{existing_output}\n{result[1]}"

[Copilot]

This changes MANA queue selection behavior: previously txonly on MANA used 8 queues, but now txonly falls back to 4 queues. If existing perf/throughput baselines depend on 8-queue txonly, this is a regression. If the intent is to use 8 queues for all MANA modes (or specifically txonly), adjust the condition accordingly or document why txonly is excluded.

Suggested change

	if self.is_mana and mode in ["rxonly", "5tswap"]:
	if self.is_mana and mode in ["rxonly", "5tswap", "txonly"]:

[mcgov]

this was a bug, this change fixes it

[Copilot]

The new 5tswap path hard-codes CPU IDs (including up to CPU 31). This makes the test fragile across VM sizes/NUMA layouts/CPU pinning and can fail even when min_core_count is met but CPU indexing/availability differs. Prefer deriving a core list from lscpu/available CPUs at runtime (and ideally keeping primary/secondary core allocations non-overlapping and NUMA-aware).

[Copilot]

The new 5tswap path hard-codes CPU IDs (including up to CPU 31). This makes the test fragile across VM sizes/NUMA layouts/CPU pinning and can fail even when min_core_count is met but CPU indexing/availability differs. Prefer deriving a core list from lscpu/available CPUs at runtime (and ideally keeping primary/secondary core allocations non-overlapping and NUMA-aware).

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

Comments suppressed due to low confidence (1)

lisa/microsoft/testsuites/dpdk/dpdkutil.py:662

• start_testpmd_concurrent now accepts Dict[DpdkTestResources, List[str]] and flattens it, but _collect_dict_result stores results in output[result[0]], which will silently overwrite earlier outputs when a kit has multiple commands. Either restrict this helper to one command per kit (validate and fail fast), or change output to collect a list of outputs per kit (or key by (kit, index)) so multi-command runs don’t lose data.

    command_pairs_as_tuples: List[Tuple[DpdkTestResources, str]] = []
    kits_and_commands = deque(node_cmd_pairs.items())
    for kit_and_commands in kits_and_commands:
        kit, commands = kit_and_commands
        for command in commands:
            command_pairs_as_tuples.append((kit, command))

    def _collect_dict_result(result: Tuple[DpdkTestResources, str]) -> None:
        output[result[0]] = result[1]

[Copilot]

In the 5tswap validation, snd_rx_pps is taken from receiver.testpmd.get_mean_rx_pps() even though this value is supposed to represent the sender’s secondary (rxonly) process. Also, the secondary process output is never parsed with process_testpmd_output, so any get_mean_rx_pps() call will be based on whichever output was parsed last (currently the sender primary txonly run), not the secondary rxonly run. Parse sender_secondary_output and compute RX PPS from that output (without overwriting the sender primary TX metrics), and use the sender’s testpmd instance rather than the receiver’s.

Suggested change

	assert_that(sender_secondary_output).described_as(
	"Sender secondary process output was empty"
	).is_not_empty()
	snd_rx_pps = receiver.testpmd.get_mean_rx_pps()
	def _get_sender_secondary_rx_pps() -> Decimal:
	rx_pps_values: List[Decimal] = []
	for line in sender_secondary_output.splitlines():
	if "rx-pps" in line.lower():
	match = re.search(r"([0-9]+(?:\.[0-9]+)?)", line)
	if match:
	rx_pps_values.append(Decimal(match.group(1)))
	assert_that(rx_pps_values).described_as(
	"Unable to parse RX PPS from sender secondary testpmd output"
	).is_not_empty()
	return sum(rx_pps_values) / Decimal(len(rx_pps_values))
	assert_that(sender_secondary_output).described_as(
	"Sender secondary process output was empty"
	).is_not_empty()
	snd_rx_pps = _get_sender_secondary_rx_pps()

[mcgov]

I have good news, they have the same output.

[Copilot]

core_list=[3, 7, 11, 17, 21] hard-codes specific CPU IDs for the sender primary process without any explanation. Since this directly controls test behavior and scheduling assumptions, please add an inline comment explaining the rationale (e.g., avoiding sibling threads/NUMA layout) or derive the core list from lscpu/NUMA topology so it adapts to different VM sizes.

[Copilot]

core_list=[1, 5, 9, 13, 19, 25, 29, 31] hard-codes CPU IDs for the sender secondary process. Please add an inline comment documenting why these exact cores were chosen (and any required minimum core count / SMT assumptions), or compute them dynamically from topology to avoid fragile pinning across SKUs.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.