On 5/31/26 7:34 PM, Nilay Shroff wrote:
On 5/28/26 8:40 AM, Geliang Tang wrote:
From: Geliang Tang<[email protected]>
Add NVMe iopolicy testing to mptcp_nvme.sh, with the default set to
"numa". It can be set to "round-robin" or "queue-depth".
Test results with 4 NVMe multipath paths and round-robin iopolicy show
that TCP and MPTCP achieve similar bandwidth:
# ./mptcp_nvme.sh tcp 4 round-robin
READ: bw=455MiB/s (478MB/s), 455MiB/s-455MiB/s (478MB/s-478MB/s),
io=4665MiB (4891MB), run=10242-10242msec
WRITE: bw=455MiB/s (477MB/s), 455MiB/s-455MiB/s (477MB/s-477MB/s),
io=4633MiB (4858MB), run=10184-10184msec
# ./mptcp_nvme.sh mptcp 4 round-robin
READ: bw=445MiB/s (466MB/s), 445MiB/s-445MiB/s (466MB/s-466MB/s),
io=4575MiB (4797MB), run=10287-10287msec
WRITE: bw=445MiB/s (467MB/s), 445MiB/s-445MiB/s (467MB/s-467MB/s),
io=4572MiB (4794MB), run=10267-10267msec
A "loss" argument is added to simulate network packet loss. When loss=1,
each veth interface is configured with "delay 5ms loss 0.5%" using tc
qdisc. Under this scenario, TCP performance is reduced by multiples
compared to MPTCP:
# ./mptcp_nvme.sh tcp 4 round-robin 1
READ: bw=144MiB/s (151MB/s), 144MiB/s-144MiB/s (151MB/s-151MB/s),
io=1909MiB (2001MB), run=13231-13231msec
WRITE: bw=100.0MiB/s (105MB/s), 100.0MiB/s-100.0MiB/s (105MB/s-105MB/s),
io=1397MiB (1465MB), run=13980-13980msec
# ./mptcp_nvme.sh mptcp 4 round-robin 1
READ: bw=428MiB/s (449MB/s), 428MiB/s-428MiB/s (449MB/s-449MB/s),
io=4524MiB (4743MB), run=10564-10564msec
WRITE: bw=431MiB/s (452MB/s), 431MiB/s-431MiB/s (452MB/s-452MB/s),
io=4513MiB (4732MB), run=10481-10481msec
These results demonstrate that MPTCP has better resilience against
packet loss compared to TCP, as it can leverage multiple subflows to
mitigate network degradation.
There are a few observations I'd like to raise:
1. It is difficult to reason about the throughput results when NVMe native
multipath is enabled together with MPTCP. In this topology, four NVMe paths
are created and the round-robin I/O policy is configured. As a result, each
I/O first goes through the NVMe multipath scheduler, which selects a path,
and is then further subjected to the MPTCP scheduler, which selects a TCP
subflow. This means there are two independent schedulers influencing I/O
placement, making it difficult to attribute the observed throughput
improvements to either NVMe multipath or MPTCP.
For throughput comparisons, it may be more meaningful to disable NVMe native
multipath (e.g., modprobe nvme_core multipath=n) when testing MPTCP. This
would
ensure that all I/O is sent through a single NVMe/TCP path while allowing
MPTCP
alone to distribute traffic across available subflows. Such a setup would
provide a clearer comparison between TCP and MPTCP.
2. The current test uses only a 128 KiB I/O size. It would be useful to include
additional I/O sizes as well, such as 4 KiB, 8 KiB, and 32 KiB, since MPTCP
and
NVMe multipath may behave differently under different workload
characteristics.
3. The fio runtime is only 10 seconds, which is relatively short for performance
evaluation. The results may be influenced by startup transients and may not
accurately reflect steady-state behavior. It would be preferable to run the
tests
for a longer duration, for example 120 seconds, to obtain more stable
measurements.
4. The tests are run on the same host by setting up veth interfaces and running
host and target under different network namespaces. It'd be useful if you
could
run this tests between real host and target systems.
One more point forgot to add:
Current tests uses symmetric path characteristics (i.e. all paths experiences
same loos or ratelimit). However it'd be useful to simulate a scenario where
paths exhibit asymmetric behavior (for instance, one path experiences loss
or increased latency compared to other). This would demonstrate the real world
network failures and it'd be interesting to see how mptcp performs compared to
native NVMe multiapth.
Thanks,
--Nilay
