Hi guys,
> > > In many usage scenarios input mbufs for rte_sched_port_enqueue() are
> > not
> > > yet in the CPU cache(s). That causes quite significant stalls due to
> > > memory
> > > latency. Current implementation tries to migitate it using SW pipeline and
> > SW
> > > prefetch techniques, but stalls are still present.
> > > Rework rte_sched_port_enqueue() to do actual fetch of all mbufs
> > metadata
> > > as a first stage of that function.
> > > That helps to minimise load stalls at further stages of enqueue() and
> > > improves overall enqueue performance.
> > > With examples/qos_sched I observed:
> > > on ICX box: up to 30% cycles reduction
> > > on CSX AND BDX: 20-15% cycles reduction
> > > I also run tests with mbufs already in the cache (one core doing RX, QOS
> > and
> > > TX).
> > > With such scenario, on all mentioned above IA boxes no performance drop
> > > was observed.
> > >
> > > Signed-off-by: Konstantin Ananyev <konstantin.anan...@intel.com>
> > > ---
> > > v2: fix clang and checkpatch complains
> > > ---
> > > lib/librte_sched/rte_sched.c | 219 +++++------------------------------
> > > 1 file changed, 31 insertions(+), 188 deletions(-)
> > >
> > > diff --git a/lib/librte_sched/rte_sched.c b/lib/librte_sched/rte_sched.c
> > index
> > > 7c5688068..41ef147e0 100644
> > > --- a/lib/librte_sched/rte_sched.c
> > > +++ b/lib/librte_sched/rte_sched.c
> > > @@ -1861,24 +1861,23 @@ debug_check_queue_slab(struct
> > > rte_sched_subport *subport, uint32_t bmp_pos, #endif /*
> > > RTE_SCHED_DEBUG */
> > >
> > > static inline struct rte_sched_subport * -rte_sched_port_subport(struct
> > > rte_sched_port *port,
> > > - struct rte_mbuf *pkt)
> > > +sched_port_subport(const struct rte_sched_port *port, struct
> > > +rte_mbuf_sched sch)
> > > {
> > > - uint32_t queue_id = rte_mbuf_sched_queue_get(pkt);
> > > + uint32_t queue_id = sch.queue_id;
> > > uint32_t subport_id = queue_id >> (port-
> > > >n_pipes_per_subport_log2 + 4);
> > >
> > > return port->subports[subport_id];
> > > }
> > >
> > > static inline uint32_t
> > > -rte_sched_port_enqueue_qptrs_prefetch0(struct rte_sched_subport
> > > *subport,
> > > - struct rte_mbuf *pkt, uint32_t subport_qmask)
> > > +sched_port_enqueue_qptrs_prefetch0(const struct rte_sched_subport
> > > *subport,
> > > + struct rte_mbuf_sched sch, uint32_t subport_qmask)
> > > {
> > > struct rte_sched_queue *q;
> > > #ifdef RTE_SCHED_COLLECT_STATS
> > > struct rte_sched_queue_extra *qe;
> > > #endif
> > > - uint32_t qindex = rte_mbuf_sched_queue_get(pkt);
> > > + uint32_t qindex = sch.queue_id;
> > > uint32_t subport_queue_id = subport_qmask & qindex;
> > >
> > > q = subport->queue + subport_queue_id; @@ -1971,197 +1970,41
> > > @@ int rte_sched_port_enqueue(struct rte_sched_port *port, struct
> > > rte_mbuf **pkts,
> > > uint32_t n_pkts)
> > > {
> > > - struct rte_mbuf *pkt00, *pkt01, *pkt10, *pkt11, *pkt20, *pkt21,
> > > - *pkt30, *pkt31, *pkt_last;
> > > - struct rte_mbuf **q00_base, **q01_base, **q10_base,
> > > **q11_base,
> > > - **q20_base, **q21_base, **q30_base, **q31_base,
> > > **q_last_base;
> > > - struct rte_sched_subport *subport00, *subport01, *subport10,
> > > *subport11,
> > > - *subport20, *subport21, *subport30, *subport31,
> > > *subport_last;
> > > - uint32_t q00, q01, q10, q11, q20, q21, q30, q31, q_last;
> > > - uint32_t r00, r01, r10, r11, r20, r21, r30, r31, r_last;
> > > - uint32_t subport_qmask;
> > > uint32_t result, i;
> > > + struct rte_mbuf_sched sch[n_pkts];
> > > + struct rte_sched_subport *subports[n_pkts];
> > > + struct rte_mbuf **q_base[n_pkts];
> > > + uint32_t q[n_pkts];
> > > +
> > > + const uint32_t subport_qmask =
> > > + (1 << (port->n_pipes_per_subport_log2 + 4)) - 1;
> > >
> > > result = 0;
> > > - subport_qmask = (1 << (port->n_pipes_per_subport_log2 + 4)) - 1;
> > >
> > > - /*
> > > - * Less then 6 input packets available, which is not enough to
> > > - * feed the pipeline
> > > - */
> > > - if (unlikely(n_pkts < 6)) {
> > > - struct rte_sched_subport *subports[5];
> > > - struct rte_mbuf **q_base[5];
> > > - uint32_t q[5];
> > > -
> > > - /* Prefetch the mbuf structure of each packet */
> > > - for (i = 0; i < n_pkts; i++)
> > > - rte_prefetch0(pkts[i]);
> > > -
> > > - /* Prefetch the subport structure for each packet */
> > > - for (i = 0; i < n_pkts; i++)
> > > - subports[i] = rte_sched_port_subport(port, pkts[i]);
> > > -
> > > - /* Prefetch the queue structure for each queue */
> > > - for (i = 0; i < n_pkts; i++)
> > > - q[i] =
> > > rte_sched_port_enqueue_qptrs_prefetch0(subports[i],
> > > - pkts[i], subport_qmask);
> > > -
> > > - /* Prefetch the write pointer location of each queue */
> > > - for (i = 0; i < n_pkts; i++) {
> > > - q_base[i] =
> > > rte_sched_subport_pipe_qbase(subports[i], q[i]);
> > > - rte_sched_port_enqueue_qwa_prefetch0(port,
> > > subports[i],
> > > - q[i], q_base[i]);
> > > - }
> > > + /* Prefetch the mbuf structure of each packet */
> > > + for (i = 0; i < n_pkts; i++)
> > > + sch[i] = pkts[i]->hash.sched;
> > >
> >
> > Hi Konstantin, thanks for the patch. In above case, all packets are touched
> > straight with any prefetch. If we consider the input burst size of 64 pkts,
> > it
> > means 512 bytes of packet addresses (8 cache-lines) which is likely to be
> > available in cache. For larger size burst, e.g. 128 or 256, there might be
> > instances when some addresses are not available the cache, may stall core.
> > How about adding explicit prefetch before starting to iterate through the
> > packets if that helps?
I don't think we need any prefetch here.
pkts[] is a sequential array, HW prefetcher should be able to do good job here.
Again in majority of use-cases pkts[] contents will already present in the
cache.
Though there is a valid concern here: n_pkts can be big, in that case we
probably
don't want to store too much on the stack and read too much from pkts[].
It is better to work in some fixed chunks (64 or so).
I can prepare v2 with these changes, if there still is an interest in this
patch.
> Exactly. Konstantin, you might not be a fan of prefetches, but the current
> enqueue implementation (as well as the dequeue) uses a prefetch
> state machine. Please keep the prefetch state machine in the scalar code.
It is not about our own preferences.
>From my measurements new version is faster and it is definitely simpler.
> Even if the examples/qos_sched might not show an advantage,
> this is just a sample app and there are some more relevant use-cases as well.
Well, I hope that examples/qos_sched reflects at least some real-world
use-cases for QOS library.
Otherwise why do we have it inside DPDK codebase?
About 'more relevant use-cases': if you do know such, can you try them with the
patch?
I would really appreciate that.
In fact, it is an ask not only to Cristian, but to all other interested parties:
if your app does use librte_sched - please try this patch and provide the
feedback.
If some tests would flag a regression - I am absolutely ok to drop the patch.
Konstantin
