[dpdk-dev,v4] app/testpmd: add option ring-bind-lcpu to bind Q with CPU
Checks
Commit Message
From: Simon Guo <wei.guo.simon@gmail.com>
Currently the rx/tx queue is allocated from the buffer pool on socket of:
- port's socket if --port-numa-config specified
- or ring-numa-config setting per port
All the above will "bind" queue to single socket per port configuration.
But it can actually archieve better performance if one port's queue can
be spread across multiple NUMA nodes, and the rx/tx queue is allocated
per lcpu socket.
This patch adds a new option "--ring-bind-lcpu"(no parameter). With
this, testpmd can utilize the PCI-e bus bandwidth on another NUMA
nodes.
When --port-numa-config or --ring-numa-config option is specified, this
--ring-bind-lcpu option will be suppressed.
Test result:
64bytes package, running in PowerPC with Mellanox
CX-4 card, single port(100G), with 8 cores, fw mode:
- Without this patch: 52.5Mpps throughput
- With this patch: 66Mpps throughput
~25% improvement
Signed-off-by: Simon Guo <wei.guo.simon@gmail.com>
---
app/test-pmd/parameters.c | 14 +++++-
app/test-pmd/testpmd.c | 112 +++++++++++++++++++++++++++++++++-------------
app/test-pmd/testpmd.h | 7 +++
3 files changed, 101 insertions(+), 32 deletions(-)
Comments
10/01/2018 08:59, wei.guo.simon@gmail.com:
> This patch adds a new option "--ring-bind-lcpu"(no parameter). With
> this, testpmd can utilize the PCI-e bus bandwidth on another NUMA
> nodes.
Waiting for a review.
Please uese --in-reply-to when sending a new version of the patch,
so we can track modifications more easily.
Thanks
Hi Simon,
> -----Original Message-----
> From: wei.guo.simon@gmail.com [mailto:wei.guo.simon@gmail.com]
> Sent: Wednesday, January 10, 2018 4:00 PM
> To: Lu, Wenzhuo <wenzhuo.lu@intel.com>
> Cc: dev@dpdk.org; Simon Guo <wei.guo.simon@gmail.com>
> Subject: [PATCH v4] app/testpmd: add option ring-bind-lcpu to bind Q with
> CPU
>
> From: Simon Guo <wei.guo.simon@gmail.com>
>
> Currently the rx/tx queue is allocated from the buffer pool on socket of:
> - port's socket if --port-numa-config specified
> - or ring-numa-config setting per port
>
> All the above will "bind" queue to single socket per port configuration.
> But it can actually archieve better performance if one port's queue can be
> spread across multiple NUMA nodes, and the rx/tx queue is allocated per
> lcpu socket.
>
> This patch adds a new option "--ring-bind-lcpu"(no parameter). With this,
> testpmd can utilize the PCI-e bus bandwidth on another NUMA nodes.
>
> When --port-numa-config or --ring-numa-config option is specified, this --
> ring-bind-lcpu option will be suppressed.
>
> Test result:
> 64bytes package, running in PowerPC with Mellanox
> CX-4 card, single port(100G), with 8 cores, fw mode:
> - Without this patch: 52.5Mpps throughput
> - With this patch: 66Mpps throughput
> ~25% improvement
>
> Signed-off-by: Simon Guo <wei.guo.simon@gmail.com>
This patch is fine. But for the new parameter, please update this doc, run_app.rst. Thanks.
Thomas,
On Mon, Jan 15, 2018 at 12:30:31AM +0100, Thomas Monjalon wrote:
> 10/01/2018 08:59, wei.guo.simon@gmail.com:
> > This patch adds a new option "--ring-bind-lcpu"(no parameter). With
> > this, testpmd can utilize the PCI-e bus bandwidth on another NUMA
> > nodes.
>
> Waiting for a review.
>
> Please uese --in-reply-to when sending a new version of the patch,
> so we can track modifications more easily.
> Thanks
Sure. Thanks for the reminder.
Thanks,
Simon
Hi Wenzuo,
On Mon, Jan 15, 2018 at 08:31:52AM +0000, Lu, Wenzhuo wrote:
> Hi Simon,
>
>
> > -----Original Message-----
> > From: wei.guo.simon@gmail.com [mailto:wei.guo.simon@gmail.com]
> > Sent: Wednesday, January 10, 2018 4:00 PM
> > To: Lu, Wenzhuo <wenzhuo.lu@intel.com>
> > Cc: dev@dpdk.org; Simon Guo <wei.guo.simon@gmail.com>
> > Subject: [PATCH v4] app/testpmd: add option ring-bind-lcpu to bind Q with
> > CPU
> >
> > From: Simon Guo <wei.guo.simon@gmail.com>
> >
> > Currently the rx/tx queue is allocated from the buffer pool on socket of:
> > - port's socket if --port-numa-config specified
> > - or ring-numa-config setting per port
> >
> > All the above will "bind" queue to single socket per port configuration.
> > But it can actually archieve better performance if one port's queue can be
> > spread across multiple NUMA nodes, and the rx/tx queue is allocated per
> > lcpu socket.
> >
> > This patch adds a new option "--ring-bind-lcpu"(no parameter). With this,
> > testpmd can utilize the PCI-e bus bandwidth on another NUMA nodes.
> >
> > When --port-numa-config or --ring-numa-config option is specified, this --
> > ring-bind-lcpu option will be suppressed.
> >
> > Test result:
> > 64bytes package, running in PowerPC with Mellanox
> > CX-4 card, single port(100G), with 8 cores, fw mode:
> > - Without this patch: 52.5Mpps throughput
> > - With this patch: 66Mpps throughput
> > ~25% improvement
> >
> > Signed-off-by: Simon Guo <wei.guo.simon@gmail.com>
> This patch is fine. But for the new parameter, please update this doc, run_app.rst. Thanks.
ah...yes. I will rework on v5.
Thanks,
- Simon
@@ -104,6 +104,10 @@
"(flag: 1 for RX; 2 for TX; 3 for RX and TX).\n");
printf(" --socket-num=N: set socket from which all memory is allocated "
"in NUMA mode.\n");
+ printf(" --ring-bind-lcpu: "
+ "specify TX/RX rings will be allocated on local socket of lcpu."
+ "It will be ignored if ring-numa-config or port-numa-config is used. "
+ "As a result, it allows one port binds to multiple NUMA nodes.\n");
printf(" --mbuf-size=N: set the data size of mbuf to N bytes.\n");
printf(" --total-num-mbufs=N: set the number of mbufs to be allocated "
"in mbuf pools.\n");
@@ -544,6 +548,7 @@
{ "interactive", 0, 0, 0 },
{ "cmdline-file", 1, 0, 0 },
{ "auto-start", 0, 0, 0 },
+ { "ring-bind-lcpu", 0, 0, 0 },
{ "eth-peers-configfile", 1, 0, 0 },
{ "eth-peer", 1, 0, 0 },
#endif
@@ -676,6 +681,10 @@
stats_period = n;
break;
}
+ if (!strcmp(lgopts[opt_idx].name, "ring-bind-lcpu")) {
+ ring_bind_lcpu |= RBL_BIND_LOCAL_MASK;
+ break;
+ }
if (!strcmp(lgopts[opt_idx].name,
"eth-peers-configfile")) {
if (init_peer_eth_addrs(optarg) != 0)
@@ -739,11 +748,14 @@
if (parse_portnuma_config(optarg))
rte_exit(EXIT_FAILURE,
"invalid port-numa configuration\n");
+ ring_bind_lcpu |= RBL_PORT_NUMA_MASK;
}
- if (!strcmp(lgopts[opt_idx].name, "ring-numa-config"))
+ if (!strcmp(lgopts[opt_idx].name, "ring-numa-config")) {
if (parse_ringnuma_config(optarg))
rte_exit(EXIT_FAILURE,
"invalid ring-numa configuration\n");
+ ring_bind_lcpu |= RBL_RING_NUMA_MASK;
+ }
if (!strcmp(lgopts[opt_idx].name, "socket-num")) {
n = atoi(optarg);
if (!new_socket_id((uint8_t)n)) {
@@ -68,6 +68,9 @@
uint8_t interactive = 0;
uint8_t auto_start = 0;
uint8_t tx_first;
+
+uint8_t ring_bind_lcpu;
+
char cmdline_filename[PATH_MAX] = {0};
/*
@@ -1410,6 +1413,43 @@ static int eth_event_callback(portid_t port_id,
return 1;
}
+static int find_local_socket(queueid_t qi, int is_rxq)
+{
+ /*
+ * try to find the local socket with following logic:
+ * 1) Find the correct stream for the queue;
+ * 2) Find the correct lcore for the stream;
+ * 3) Find the correct socket for the lcore;
+ */
+ int i, j, socket = NUMA_NO_CONFIG;
+
+ /* find the stream based on queue no*/
+ for (i = 0; i < nb_fwd_streams; i++) {
+ if (is_rxq) {
+ if (fwd_streams[i]->rx_queue == qi)
+ break;
+ } else {
+ if (fwd_streams[i]->tx_queue == qi)
+ break;
+ }
+ }
+ if (i == nb_fwd_streams)
+ return NUMA_NO_CONFIG;
+
+ /* find the lcore based on stream idx */
+ for (j = 0; j < nb_lcores; j++) {
+ if (fwd_lcores[j]->stream_idx == i)
+ break;
+ }
+ if (j == nb_lcores)
+ return NUMA_NO_CONFIG;
+
+ /* find the socket for the lcore */
+ socket = rte_lcore_to_socket_id(fwd_lcores_cpuids[j]);
+
+ return socket;
+}
+
int
start_port(portid_t pid)
{
@@ -1469,14 +1509,18 @@ static int eth_event_callback(portid_t port_id,
port->need_reconfig_queues = 0;
/* setup tx queues */
for (qi = 0; qi < nb_txq; qi++) {
+ int socket = port->socket_id;
if ((numa_support) &&
(txring_numa[pi] != NUMA_NO_CONFIG))
- diag = rte_eth_tx_queue_setup(pi, qi,
- nb_txd,txring_numa[pi],
- &(port->tx_conf));
- else
- diag = rte_eth_tx_queue_setup(pi, qi,
- nb_txd,port->socket_id,
+ socket = txring_numa[pi];
+ else if (ring_bind_lcpu) {
+ int ret = find_local_socket(qi, 0);
+ if (ret != NUMA_NO_CONFIG)
+ socket = ret;
+ }
+
+ diag = rte_eth_tx_queue_setup(pi, qi,
+ nb_txd, socket,
&(port->tx_conf));
if (diag == 0)
@@ -1495,35 +1539,28 @@ static int eth_event_callback(portid_t port_id,
}
/* setup rx queues */
for (qi = 0; qi < nb_rxq; qi++) {
+ int socket = port->socket_id;
if ((numa_support) &&
- (rxring_numa[pi] != NUMA_NO_CONFIG)) {
- struct rte_mempool * mp =
- mbuf_pool_find(rxring_numa[pi]);
- if (mp == NULL) {
- printf("Failed to setup RX queue:"
- "No mempool allocation"
- " on the socket %d\n",
- rxring_numa[pi]);
- return -1;
- }
-
- diag = rte_eth_rx_queue_setup(pi, qi,
- nb_rxd,rxring_numa[pi],
- &(port->rx_conf),mp);
- } else {
- struct rte_mempool *mp =
- mbuf_pool_find(port->socket_id);
- if (mp == NULL) {
- printf("Failed to setup RX queue:"
+ (rxring_numa[pi] != NUMA_NO_CONFIG))
+ socket = rxring_numa[pi];
+ else if (ring_bind_lcpu) {
+ int ret = find_local_socket(qi, 1);
+ if (ret != NUMA_NO_CONFIG)
+ socket = ret;
+ }
+
+ struct rte_mempool *mp =
+ mbuf_pool_find(socket);
+ if (mp == NULL) {
+ printf("Failed to setup RX queue:"
"No mempool allocation"
" on the socket %d\n",
- port->socket_id);
- return -1;
- }
- diag = rte_eth_rx_queue_setup(pi, qi,
- nb_rxd,port->socket_id,
- &(port->rx_conf), mp);
+ socket);
+ return -1;
}
+ diag = rte_eth_rx_queue_setup(pi, qi,
+ nb_rxd, socket,
+ &(port->rx_conf), mp);
if (diag == 0)
continue;
@@ -2414,6 +2451,19 @@ uint8_t port_is_bonding_slave(portid_t slave_pid)
"but nb_txq=%d will prevent to fully test it.\n",
nb_rxq, nb_txq);
+ if (ring_bind_lcpu & RBL_BIND_LOCAL_MASK) {
+ if (ring_bind_lcpu &
+ (RBL_RING_NUMA_MASK | RBL_PORT_NUMA_MASK)) {
+ printf("ring-bind-lcpu option is suppressed by "
+ "ring-numa-config or port-numa-config option\n");
+ ring_bind_lcpu = 0;
+ } else {
+ printf("RingBuffer bind with local CPU selected\n");
+ ring_bind_lcpu = 1;
+ }
+ } else
+ ring_bind_lcpu = 0;
+
init_config();
if (start_port(RTE_PORT_ALL) != 0)
rte_exit(EXIT_FAILURE, "Start ports failed\n");
@@ -328,6 +328,13 @@ struct queue_stats_mappings {
extern uint8_t interactive;
extern uint8_t auto_start;
extern uint8_t tx_first;
+
+/* for --ring-bind-lcpu option checking, RBL means Ring Bind Lcpu related */
+#define RBL_BIND_LOCAL_MASK (1 << 0)
+#define RBL_RING_NUMA_MASK (1 << 1)
+#define RBL_PORT_NUMA_MASK (1 << 2)
+extern uint8_t ring_bind_lcpu;
+
extern char cmdline_filename[PATH_MAX]; /**< offline commands file */
extern uint8_t numa_support; /**< set by "--numa" parameter */
extern uint16_t port_topology; /**< set by "--port-topology" parameter */