[lustre-devel] [PATCH 09/24] lnet: o2iblnd: Salt comp_vector
James Simmons
jsimmons at infradead.org
Mon Sep 5 18:55:22 PDT 2022
From: Ian Ziemba <ian.ziemba at hpe.com>
If conns_per_peer is greater than 1, all the connections targeting
the same peer are assigned the same comp_vector. This results in
multiple IB CQs targeting the same peer to be serialized on a single
comp_vector.
Help spread out the IB CQ work to multiple cores by salting
comp_vector based on number of connections.
1 client to 1 server LST 1M write results with 4 conns_per_peer and
RXE configured to spread out work based on comp_vector.
Before: 1377.92 MB/s
After: 3828.48 MB/s
HPE-bug-id: LUS-11043
WC-bug-id: https://jira.whamcloud.com/browse/LU-16078
Lustre-commit: 1ef1fa06b20c424f5 ("LU-16078 o2iblnd: Salt comp_vector")
Signed-off-by: Ian Ziemba <ian.ziemba at hpe.com>
Reviewed-on: https://review.whamcloud.com/48148
Reviewed-by: Andreas Dilger <adilger at whamcloud.com>
Reviewed-by: Oleg Drokin <green at whamcloud.com>
Signed-off-by: James Simmons <jsimmons at infradead.org>
---
net/lnet/klnds/o2iblnd/o2iblnd.c | 14 +++++++++++---
net/lnet/klnds/o2iblnd/o2iblnd.h | 2 ++
2 files changed, 13 insertions(+), 3 deletions(-)
diff --git a/net/lnet/klnds/o2iblnd/o2iblnd.c b/net/lnet/klnds/o2iblnd/o2iblnd.c
index ea28c65..c713528 100644
--- a/net/lnet/klnds/o2iblnd/o2iblnd.c
+++ b/net/lnet/klnds/o2iblnd/o2iblnd.c
@@ -338,6 +338,7 @@ int kiblnd_create_peer(struct lnet_ni *ni, struct kib_peer_ni **peerp,
peer_ni->ibp_queue_depth = ni->ni_net->net_tunables.lct_peer_tx_credits;
peer_ni->ibp_queue_depth_mod = 0; /* try to use the default */
kref_init(&peer_ni->ibp_kref);
+ atomic_set(&peer_ni->ibp_nconns, 0);
INIT_HLIST_NODE(&peer_ni->ibp_list);
INIT_LIST_HEAD(&peer_ni->ibp_conns);
@@ -569,7 +570,7 @@ static int kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
int vectors;
int off;
int i;
- lnet_nid_t nid = conn->ibc_peer->ibp_nid;
+ lnet_nid_t ibp_nid;
vectors = conn->ibc_cmid->device->num_comp_vectors;
if (vectors <= 1)
@@ -579,8 +580,13 @@ static int kiblnd_get_completion_vector(struct kib_conn *conn, int cpt)
if (!mask)
return 0;
- /* hash NID to CPU id in this partition... */
- off = do_div(nid, cpumask_weight(*mask));
+ /* hash NID to CPU id in this partition... when targeting a single peer
+ * with multiple QPs, to engage more cores in CQ processing to a single
+ * peer, use ibp_nconns to salt the value the comp_vector value
+ */
+ ibp_nid = conn->ibc_peer->ibp_nid +
+ atomic_read(&conn->ibc_peer->ibp_nconns);
+ off = do_div(ibp_nid, cpumask_weight(*mask));
for_each_cpu(i, *mask) {
if (!off--)
return i % vectors;
@@ -889,6 +895,7 @@ struct kib_conn *kiblnd_create_conn(struct kib_peer_ni *peer_ni,
conn->ibc_state = state;
/* 1 more conn */
+ atomic_inc(&peer_ni->ibp_nconns);
atomic_inc(&net->ibn_nconns);
return conn;
@@ -954,6 +961,7 @@ void kiblnd_destroy_conn(struct kib_conn *conn)
kiblnd_peer_decref(peer_ni);
rdma_destroy_id(cmid);
+ atomic_dec(&peer_ni->ibp_nconns);
atomic_dec(&net->ibn_nconns);
}
}
diff --git a/net/lnet/klnds/o2iblnd/o2iblnd.h b/net/lnet/klnds/o2iblnd/o2iblnd.h
index 0066e85..56d486f 100644
--- a/net/lnet/klnds/o2iblnd/o2iblnd.h
+++ b/net/lnet/klnds/o2iblnd/o2iblnd.h
@@ -522,6 +522,8 @@ struct kib_peer_ni {
u16 ibp_queue_depth;
/* reduced value which allows conn to be created if max fails */
u16 ibp_queue_depth_mod;
+ /* Number of connections allocated. */
+ atomic_t ibp_nconns;
};
extern struct kib_data kiblnd_data;
--
1.8.3.1
More information about the lustre-devel
mailing list