[PATCH v1 0/8] event/slb: introducing software load balancer eventdev PMD
Sarkar, Tirthendu
tirthendu.sarkar at intel.com
Wed Nov 11 15:22:03 CET 2026
On Mon, Nov 9, 2026 at 6:25 PM Jerin Jacob <jerinjacobk at gmail.com> wrote:
> Lets get the performance numbers for SLB vs DSW based on same test
> case so that we can decide.
> Performance and scalability will be the aspect to decide the usage of
> YET ANOTHER SW driver.
Hi Pravin,
A few points on top of Stephen's review, plus some early numbers since
Jerin asked for SLB vs DSW data.
+1 to Stephen's two Errors in patch 2 (reorder slot reuse when RELEASE
returns credits before the slot drains, and the lost wakeup in
slb_reorder_drain()) and to the credit-threshold issue in patch 1.
Those look like the ones that can lose or strand events.
Additional comments:
1. slb_event.c, slb_event_enqueue_burst_generic(): the return value of
rte_event_ring_enqueue_burst() is ignored on every path and the
function always returns events_len. Correctness then depends
entirely on credits guaranteeing ring space. If that ever fails
(e.g. the credit-threshold or reorder issues above, or application
misuse), events are silently dropped and their credits leak. At
minimum, check the return value and report a short enqueue.
Because the burst is sorted by destination queue before enqueue,
a clean prefix count isn't possible after a partial failure, so
this probably needs a space check up front or an explicit design
statement that rings can never fill.
2. RTE_VERIFY(num_forward + num_release <= cur_pending) in the
enqueue fast path aborts the whole process on application misuse
(forwarding more events than were dequeued on that port). Please
fail the enqueue and set rte_errno instead.
3. slb_configure(): max_inflight = RTE_MAX(nb_events_limit,
nb_ports * SLB_PORT_MAX_CREDITS), so the device can hold more
events than the application configured via nb_events_limit. If
that is intended (to avoid credit-cache deadlock), please document
it in slb.rst.
4. slb_queue_setup() ignores queue_conf->priority; priority comes only
from the link priority, yet dev_info reports
max_event_queue_priority_levels = 2. Either honour queue priority or
report 1 and document that priority is per link.
Performance
-----------
Preliminary data from dpdk-test-eventdev perf_queue, series applied on
v26.11-rc2, release build, Xeon 6505P (2x12C), physical cores only,
turbo on, default device settings, worker_deq_depth 16, 1024 flows.
Mpps, 1 synthetic producer, workers on socket 0:
stlist driver 1w 2w 4w 8w
p slb 79.7 71.5 44.5 34.0
p dsw 50.1 46.3 45.1 41.9
a slb 54.3 65.3 57.0 43.9
a dsw 50.3 47.3 46.7 40.5
a,a,a slb 20.8 16.9 22.9 26.7
a,a,a dsw 20.8 18.4 20.1 18.2
o slb 61.7 29.5 36.6 27.4
o sw 7.7 10.0 11.7 13.7
o,a slb 36.0 19.5 28.3 28.9
o,a sw 3.8 5.7 7.5 7.2
(event_dsw does not support ordered queues, so ordered rows are
against event_sw.)
Observations:
- SLB is ahead at low worker counts but loses throughput faster as
workers are added on single-stage parallel/atomic queues, which is
consistent with all workers contending on one MP/MC ring per queue.
- Ordered queues are well ahead of event_sw (the only other SW PMD with
ordered support), but SLB ordered throughput halves going from 1 to
2 workers (61.7 -> 29.5 Mpps, reproduced), so the "close to
parallel" claim only holds for a single worker (parallel: 79.7 ->
71.5). This looks like the serialized reorder drain.
- With workers spanning both sockets (16/20 workers) all drivers drop
sharply; SLB falls to 5-10 Mpps vs DSW 3-14 depending on stlist.
- Note that with multiple producers perf_queue creates a separate
queue set per producer, which splits SLB's per-queue contention; SLB
scales very well in that mode (4 producers, 8 workers, atomic:
~235 Mpps vs ~158 for DSW at worker_deq_depth 16), but that does not
reflect a pipeline where all traffic shares one atomic queue.
- order_queue passes with up to 16 workers. In that test SLB slows down
as workers are added: with 8 workers it sustained ~1.0 M events/s vs
~2.5 M events/s for event_sw (steady-state rate from the progress
counter, 20M events); with 1-2 workers SLB was faster than event_sw
(rough end-to-end timings, ~3.2 vs ~1.6-2.1 M events/s).
These are untuned runs. In the 1-producer table the SLB/DSW atomic and
a,a,a rows, DSW parallel and SLB ordered were run twice and agreed
within ~5%; the other rows are single runs. Treat them as indicative.
It would help if v2 included your own perf_queue/pipeline numbers with
the exact command lines, ideally including a shared-queue case and
workers on more than one socket.
Thanks,
Tirthendu
More information about the dev
mailing list