Thanks for the patch, the implementation looks good to me too.
During testing I kept noticing that it's way too easy to make OVS
completely unresponsive. As you point out in the documentation by having
dpdk-lcore-mask the same as pmd-cpu-mask, OVS cannot even be killed (a kill
-9 is required). I wonder what happens if one tries to set pmd-cpu-mask to
every core in the system.
As a way to mitigate the risk perhaps we can avoid setting the main thread
affinity to the first core in dpdk-lcore-mask by _always_ restoring it in
dpdk_init__(), also if auto_determine is false.
Perhaps we should start explicitly prohibiting creating a pmd thread on the
first core in dpdk-lcore-mask (I get why previous version of this didn't do
it on core 0. Perhaps we can generalize that to the first core in
dpdk-lcore-mask).
What's the behavior of other DPDK applications?
Thanks,
Daniele
2016-07-27 5:28 GMT-07:00 Kavanagh, Mark B <mark.b.kavan...@intel.com>:
> >
> >Set the DPDK pmd thread scheduling policy to SCHED_RR and static
> >priority to highest priority value of the policy. This is to deal with
> >pmd thread starvation case where another cpu hogging process can get
> >scheduled/affinitized on to the same core the pmd thread is running
> >there by significantly impacting the datapath performance.
> >
> >Setting the realtime scheduling policy to the pmd threads is one step
> >towards Fastpath Service Assurance in OVS DPDK.
> >
> >The realtime scheduling policy is applied only when CPU mask is passed
> >to 'pmd-cpu-mask'. For example:
> >
> > * In the absence of pmd-cpu-mask, one pmd thread shall be created
> > and default scheduling policy and priority gets applied.
> >
> > * If pmd-cpu-mask is specified, one or more pmd threads shall be
> > spawned on the corresponding core(s) in the mask and real time
> > scheduling policy SCHED_RR and highest priority of the policy is
> > applied to the pmd thread(s).
> >
> >To reproduce the pmd thread starvation case:
> >
> >ovs-vsctl set Open_vSwitch . other_config:pmd-cpu-mask=6
> >taskset 0x2 cat /dev/zero > /dev/null &
> >
> >With this commit, it is recommended that the OVS control thread and pmd
> >thread shouldn't be pinned to same core ('dpdk-lcore-mask','pmd-cpu-mask'
> >should be non-overlapping). Also other processes with same affinity as
> >PMD thread will be unresponsive.
> >
> >Signed-off-by: Bhanuprakash Bodireddy <bhanuprakash.bodire...@intel.com>
>
> LGTM - Acked-by: mark.b.kavan...@intel.com
>
> >---
> >v4->v5:
> >* Reword Note section in DPDK-ADVANCED.md
> >
> >v3->v4:
> >* Document update
> >* Use ovs_strerror for reporting errors in lib-numa.c
> >
> >v2->v3:
> >* Move set_priority() function to lib/ovs-numa.c
> >* Apply realtime scheduling policy and priority to pmd thread only if
> > pmd-cpu-mask is passed.
> >* Update INSTALL.DPDK-ADVANCED.
> >
> >v1->v2:
> >* Removed #ifdef and introduced dummy function "pmd_thread_setpriority"
> > in netdev-dpdk.h
> >* Rebase
> >
> > INSTALL.DPDK-ADVANCED.md | 17 +++++++++++++----
> > lib/dpif-netdev.c | 9 +++++++++
> > lib/ovs-numa.c | 18 ++++++++++++++++++
> > lib/ovs-numa.h | 1 +
> > 4 files changed, 41 insertions(+), 4 deletions(-)
> >
> >diff --git a/INSTALL.DPDK-ADVANCED.md b/INSTALL.DPDK-ADVANCED.md
> >index 9ae536d..d76cb4e 100644
> >--- a/INSTALL.DPDK-ADVANCED.md
> >+++ b/INSTALL.DPDK-ADVANCED.md
> >@@ -205,8 +205,10 @@ needs to be affinitized accordingly.
> > pmd thread is CPU bound, and needs to be affinitized to isolated
> > cores for optimum performance.
> >
> >- By setting a bit in the mask, a pmd thread is created and pinned
> >- to the corresponding CPU core. e.g. to run a pmd thread on core 2
> >+ By setting a bit in the mask, a pmd thread is created, pinned
> >+ to the corresponding CPU core and the scheduling policy SCHED_RR
> >+ along with maximum priority of the policy applied to the pmd thread.
> >+ e.g. to pin a pmd thread on core 2
> >
> > `ovs-vsctl set Open_vSwitch . other_config:pmd-cpu-mask=4`
> >
> >@@ -234,8 +236,10 @@ needs to be affinitized accordingly.
> > responsible for different ports/rxq's. Assignment of ports/rxq's to
> > pmd threads is done automatically.
> >
> >- A set bit in the mask means a pmd thread is created and pinned
> >- to the corresponding CPU core. e.g. to run pmd threads on core 1 and 2
> >+ A set bit in the mask means a pmd thread is created, pinned to the
> >+ corresponding CPU core and the scheduling policy SCHED_RR with highest
> >+ priority of the scheduling policy applied to pmd thread.
> >+ e.g. to run pmd threads on core 1 and 2
> >
> > `ovs-vsctl set Open_vSwitch . other_config:pmd-cpu-mask=6`
> >
> >@@ -246,6 +250,11 @@ needs to be affinitized accordingly.
> >
> > NIC port0 <-> OVS <-> VM <-> OVS <-> NIC port 1
> >
> >+ Note: It is recommended that the OVS control thread and pmd thread
> >+ shouldn't be pinned to same core i.e 'dpdk-lcore-mask' and
> 'pmd-cpu-mask'
> >+ cpu mask settings should be non-overlapping. Also other processes with
> >+ same affinity as PMD thread will be unresponsive.
> >+
> > ### 4.3 DPDK physical port Rx Queues
> >
> > `ovs-vsctl set Interface <DPDK interface> options:n_rxq=<integer>`
> >diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
> >index f05ca4e..b85600b 100644
> >--- a/lib/dpif-netdev.c
> >+++ b/lib/dpif-netdev.c
> >@@ -2841,6 +2841,15 @@ pmd_thread_main(void *f_)
> > ovs_numa_thread_setaffinity_core(pmd->core_id);
> > dpdk_set_lcore_id(pmd->core_id);
> > poll_cnt = pmd_load_queues_and_ports(pmd, &poll_list);
> >+
> >+ /* When cpu affinity mask explicitly set using pmd-cpu-mask, pmd
> thread's
> >+ * scheduling policy is set to SCHED_RR and the priority to highest
> priority
> >+ * of SCHED_RR policy. In the absence of pmd-cpu-mask, default
> scheduling
> >+ * policy and priority shall apply to pmd thread.
> >+ */
> >+ if (pmd->dp->pmd_cmask) {
> >+ ovs_numa_thread_setpriority(SCHED_RR);
> >+ }
> > reload:
> > emc_cache_init(&pmd->flow_cache);
> >
> >diff --git a/lib/ovs-numa.c b/lib/ovs-numa.c
> >index c8173e0..428f274 100644
> >--- a/lib/ovs-numa.c
> >+++ b/lib/ovs-numa.c
> >@@ -613,3 +613,21 @@ int ovs_numa_thread_setaffinity_core(unsigned
> core_id OVS_UNUSED)
> > return EOPNOTSUPP;
> > #endif /* __linux__ */
> > }
> >+
> >+void
> >+ovs_numa_thread_setpriority(int policy)
> >+{
> >+ if (dummy_numa) {
> >+ return;
> >+ }
> >+
> >+ struct sched_param threadparam;
> >+ int err;
> >+
> >+ memset(&threadparam, 0, sizeof(threadparam));
> >+ threadparam.sched_priority = sched_get_priority_max(policy);
> >+ err = pthread_setschedparam(pthread_self(), policy, &threadparam);
> >+ if (err) {
> >+ VLOG_ERR("Thread priority error %s",ovs_strerror(err));
> >+ }
> >+}
> >diff --git a/lib/ovs-numa.h b/lib/ovs-numa.h
> >index be836b2..94f0884 100644
> >--- a/lib/ovs-numa.h
> >+++ b/lib/ovs-numa.h
> >@@ -56,6 +56,7 @@ void ovs_numa_unpin_core(unsigned core_id);
> > struct ovs_numa_dump *ovs_numa_dump_cores_on_numa(int numa_id);
> > void ovs_numa_dump_destroy(struct ovs_numa_dump *);
> > int ovs_numa_thread_setaffinity_core(unsigned core_id);
> >+void ovs_numa_thread_setpriority(int policy);
> >
> > #define FOR_EACH_CORE_ON_NUMA(ITER, DUMP) \
> > LIST_FOR_EACH((ITER), list_node, &(DUMP)->dump)
> >--
> >2.4.11
>
>
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
