| Message ID | 1512418228-83914-1-git-send-email-bhanuprakash.bodireddy@intel.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [ovs-dev,v2] packets: Prefetch the packet metadata in cacheline1. | expand |
Tested on Intel Haswell platform and Acked! (Provided there are no performance regression are reported so far on other platforms for various tests too) Regards _Sugesh > -----Original Message----- > From: Bodireddy, Bhanuprakash > Sent: Monday, December 4, 2017 8:10 PM > To: dev@openvswitch.org > Cc: Chandran, Sugesh <sugesh.chandran@intel.com>; Bodireddy, Bhanuprakash > <bhanuprakash.bodireddy@intel.com>; Ben Pfaff <blp@ovn.org> > Subject: [PATCH v2] packets: Prefetch the packet metadata in cacheline1. > > pkt_metadata_prefetch_init() is used to prefetch the packet metadata before > initializing the metadata in pkt_metadata_init(). This is done for every packet in > userspace datapath and is performance critical. > > Commit 99fc16c0 prefetches only cachline0 and cacheline2 as the metadata > part of respective cachelines will be initialized by pkt_metadata_init(). > > However in VXLAN case when popping the vxlan header, > netdev_vxlan_pop_header() invokes pkt_metadata_init_tnl() which zeroes out > metadata part of > cacheline1 that wasn't prefetched earlier and causes performance degradation. > > By prefetching cacheline1, 9% performance improvement is observed with vxlan > decapsulation test case for packet sizes of 118 bytes. Performance variation is > observed based on CFLAGS. > > CFLAGS="-O2" CFLAGS="-O2 -msse4.2" > Master 4.667 Mpps Master 4.710 Mpps > With Patch 5.045 Mpps With Patch 5.097 Mpps > > CFLAGS="-O2 -march=native" CFLAGS="-Ofast -march=native" > Master 5.072 Mpps Master 5.349 Mpps > With Patch 5.193 Mpps With Patch 5.378 Mpps > > CC: Ben Pfaff <blp@ovn.org> > Fixes: 99fc16c0 ("Reorganize the pkt_metadata structure.") > Signed-off-by: Bhanuprakash Bodireddy <bhanuprakash.bodireddy@intel.com> > --- > v2->v1 > * Include the throughput stats with different CFLAG options. > > lib/packets.h | 7 ++++++- > 1 file changed, 6 insertions(+), 1 deletion(-) > > diff --git a/lib/packets.h b/lib/packets.h index 13ea46d..74bec5d 100644 > --- a/lib/packets.h > +++ b/lib/packets.h > @@ -159,7 +159,8 @@ pkt_metadata_init(struct pkt_metadata *md, > odp_port_t port) } > > /* This function prefetches the cachelines touched by pkt_metadata_init() > - * For performance reasons the two functions should be kept in sync. */ > + * and pkt_metadata_init_tnl(). For performance reasons the two > + functions > + * should be kep in sync. */ > static inline void > pkt_metadata_prefetch_init(struct pkt_metadata *md) { @@ -167,6 +168,10 > @@ pkt_metadata_prefetch_init(struct pkt_metadata *md) > * be initialized later in pkt_metadata_init(). */ > OVS_PREFETCH(md->cacheline0); > > + /* Prefetch cacheline1 as members of this cacheline will be zeroed out > + * in pkt_metadata_init_tnl(). */ > + OVS_PREFETCH(md->cacheline1); > + > /* Prefetch cachline2 as ip_dst & ipv6_dst fields will be initialized. */ > OVS_PREFETCH(md->cacheline2); > } > -- > 2.4.11
diff --git a/lib/packets.h b/lib/packets.h index 13ea46d..74bec5d 100644 --- a/lib/packets.h +++ b/lib/packets.h @@ -159,7 +159,8 @@ pkt_metadata_init(struct pkt_metadata *md, odp_port_t port) } /* This function prefetches the cachelines touched by pkt_metadata_init() - * For performance reasons the two functions should be kept in sync. */ + * and pkt_metadata_init_tnl(). For performance reasons the two functions + * should be kep in sync. */ static inline void pkt_metadata_prefetch_init(struct pkt_metadata *md) { @@ -167,6 +168,10 @@ pkt_metadata_prefetch_init(struct pkt_metadata *md) * be initialized later in pkt_metadata_init(). */ OVS_PREFETCH(md->cacheline0); + /* Prefetch cacheline1 as members of this cacheline will be zeroed out + * in pkt_metadata_init_tnl(). */ + OVS_PREFETCH(md->cacheline1); + /* Prefetch cachline2 as ip_dst & ipv6_dst fields will be initialized. */ OVS_PREFETCH(md->cacheline2); }
pkt_metadata_prefetch_init() is used to prefetch the packet metadata before initializing the metadata in pkt_metadata_init(). This is done for every packet in userspace datapath and is performance critical. Commit 99fc16c0 prefetches only cachline0 and cacheline2 as the metadata part of respective cachelines will be initialized by pkt_metadata_init(). However in VXLAN case when popping the vxlan header, netdev_vxlan_pop_header() invokes pkt_metadata_init_tnl() which zeroes out metadata part of cacheline1 that wasn't prefetched earlier and causes performance degradation. By prefetching cacheline1, 9% performance improvement is observed with vxlan decapsulation test case for packet sizes of 118 bytes. Performance variation is observed based on CFLAGS. CFLAGS="-O2" CFLAGS="-O2 -msse4.2" Master 4.667 Mpps Master 4.710 Mpps With Patch 5.045 Mpps With Patch 5.097 Mpps CFLAGS="-O2 -march=native" CFLAGS="-Ofast -march=native" Master 5.072 Mpps Master 5.349 Mpps With Patch 5.193 Mpps With Patch 5.378 Mpps CC: Ben Pfaff <blp@ovn.org> Fixes: 99fc16c0 ("Reorganize the pkt_metadata structure.") Signed-off-by: Bhanuprakash Bodireddy <bhanuprakash.bodireddy@intel.com> --- v2->v1 * Include the throughput stats with different CFLAG options. lib/packets.h | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)