[v2] app/testpmd: tx pkt clones parameter in flowgen
Checks
Commit Message
When testing high performance numbers, it is often that CPU performance
limits the max values device can reach (both in pps and in gbps)
Here instead of recreating each packet separately, we use clones counter
to resend the same mbuf to the line multiple times.
PMDs handle that transparently due to reference counting inside of mbuf.
Reaching max PPS on small packet sizes helps here:
Some data from our 2 port x 50G device. Using 2*6 tx queues, 64b packets,
PowerEdge R7525, AMD EPYC 7452:
./build/app/dpdk-testpmd -l 32-63 -- --forward-mode=flowgen \
--rxq=6 --txq=6 --disable-crc-strip --burst=512 \
--flowgen-clones=0 --txd=4096 --stats-period=1 --txpkts=64
Gives ~46MPPS TX output:
Tx-pps: 22926849 Tx-bps: 11738590176
Tx-pps: 23642629 Tx-bps: 12105024112
Setting flowgen-clones to 512 pushes TX almost to our device
physical limit (68MPPS) using same 2*6 queues(cores):
Tx-pps: 34357556 Tx-bps: 17591073696
Tx-pps: 34353211 Tx-bps: 17588802640
Doing similar measurements per core, I see one core can do
6.9MPPS (without clones) vs 11MPPS (with clones)
Verified on Marvell qede and atlantic PMDs.
v2:
- fixed warning on uninit var
v1:
- fixes on Ferruh's comments
rfc v2: http://patchwork.dpdk.org/patch/78800/
- increment ref counter for each mbuf pointer copy
rfc v1: http://patchwork.dpdk.org/patch/78674/
Signed-off-by: Igor Russkikh <irusskikh@marvell.com>
---
app/test-pmd/flowgen.c | 107 ++++++++++++++------------
app/test-pmd/parameters.c | 10 +++
app/test-pmd/testpmd.c | 1 +
app/test-pmd/testpmd.h | 1 +
doc/guides/testpmd_app_ug/run_app.rst | 7 ++
5 files changed, 78 insertions(+), 48 deletions(-)
Comments
On 1/21/2021 6:05 PM, Igor Russkikh wrote:
> When testing high performance numbers, it is often that CPU performance
> limits the max values device can reach (both in pps and in gbps)
>
> Here instead of recreating each packet separately, we use clones counter
> to resend the same mbuf to the line multiple times.
>
> PMDs handle that transparently due to reference counting inside of mbuf.
>
> Reaching max PPS on small packet sizes helps here:
> Some data from our 2 port x 50G device. Using 2*6 tx queues, 64b packets,
> PowerEdge R7525, AMD EPYC 7452:
>
> ./build/app/dpdk-testpmd -l 32-63 -- --forward-mode=flowgen \
> --rxq=6 --txq=6 --disable-crc-strip --burst=512 \
> --flowgen-clones=0 --txd=4096 --stats-period=1 --txpkts=64
>
> Gives ~46MPPS TX output:
>
> Tx-pps: 22926849 Tx-bps: 11738590176
> Tx-pps: 23642629 Tx-bps: 12105024112
>
> Setting flowgen-clones to 512 pushes TX almost to our device
> physical limit (68MPPS) using same 2*6 queues(cores):
>
> Tx-pps: 34357556 Tx-bps: 17591073696
> Tx-pps: 34353211 Tx-bps: 17588802640
>
> Doing similar measurements per core, I see one core can do
> 6.9MPPS (without clones) vs 11MPPS (with clones)
>
> Verified on Marvell qede and atlantic PMDs.
>
> v2:
> - fixed warning on uninit var
> v1:
> - fixes on Ferruh's comments
>
> rfc v2: http://patchwork.dpdk.org/patch/78800/
> - increment ref counter for each mbuf pointer copy
> rfc v1: http://patchwork.dpdk.org/patch/78674/
>
> Signed-off-by: Igor Russkikh <irusskikh@marvell.com>
Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
Also a testpmd command to set the 'flowgen-clones' can be useful, for
interactive mode. If you have time for it, can you please work on it too?
Applied to dpdk-next-net/main, thanks.
> > Signed-off-by: Igor Russkikh <irusskikh@marvell.com>
>
> Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>
>
> Also a testpmd command to set the 'flowgen-clones' can be useful, for
> interactive mode. If you have time for it, can you please work on it too?
>
>
> Applied to dpdk-next-net/main, thanks.
Thanks, Ferruh.
Good idea, will work.
Igor
@@ -85,7 +85,7 @@ pkt_burst_flow_gen(struct fwd_stream *fs)
unsigned pkt_size = tx_pkt_length - 4; /* Adjust FCS */
struct rte_mbuf *pkts_burst[MAX_PKT_BURST];
struct rte_mempool *mbp;
- struct rte_mbuf *pkt;
+ struct rte_mbuf *pkt = NULL;
struct rte_ether_hdr *eth_hdr;
struct rte_ipv4_hdr *ip_hdr;
struct rte_udp_hdr *udp_hdr;
@@ -94,6 +94,7 @@ pkt_burst_flow_gen(struct fwd_stream *fs)
uint16_t nb_rx;
uint16_t nb_tx;
uint16_t nb_pkt;
+ uint16_t nb_clones = nb_pkt_flowgen_clones;
uint16_t i;
uint32_t retry;
uint64_t tx_offloads;
@@ -123,53 +124,63 @@ pkt_burst_flow_gen(struct fwd_stream *fs)
ol_flags |= PKT_TX_MACSEC;
for (nb_pkt = 0; nb_pkt < nb_pkt_per_burst; nb_pkt++) {
- pkt = rte_mbuf_raw_alloc(mbp);
- if (!pkt)
- break;
-
- pkt->data_len = pkt_size;
- pkt->next = NULL;
-
- /* Initialize Ethernet header. */
- eth_hdr = rte_pktmbuf_mtod(pkt, struct rte_ether_hdr *);
- rte_ether_addr_copy(&cfg_ether_dst, ð_hdr->d_addr);
- rte_ether_addr_copy(&cfg_ether_src, ð_hdr->s_addr);
- eth_hdr->ether_type = rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV4);
-
- /* Initialize IP header. */
- ip_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
- memset(ip_hdr, 0, sizeof(*ip_hdr));
- ip_hdr->version_ihl = RTE_IPV4_VHL_DEF;
- ip_hdr->type_of_service = 0;
- ip_hdr->fragment_offset = 0;
- ip_hdr->time_to_live = IP_DEFTTL;
- ip_hdr->next_proto_id = IPPROTO_UDP;
- ip_hdr->packet_id = 0;
- ip_hdr->src_addr = rte_cpu_to_be_32(cfg_ip_src);
- ip_hdr->dst_addr = rte_cpu_to_be_32(cfg_ip_dst +
- next_flow);
- ip_hdr->total_length = RTE_CPU_TO_BE_16(pkt_size -
- sizeof(*eth_hdr));
- ip_hdr->hdr_checksum = ip_sum((unaligned_uint16_t *)ip_hdr,
- sizeof(*ip_hdr));
-
- /* Initialize UDP header. */
- udp_hdr = (struct rte_udp_hdr *)(ip_hdr + 1);
- udp_hdr->src_port = rte_cpu_to_be_16(cfg_udp_src);
- udp_hdr->dst_port = rte_cpu_to_be_16(cfg_udp_dst);
- udp_hdr->dgram_cksum = 0; /* No UDP checksum. */
- udp_hdr->dgram_len = RTE_CPU_TO_BE_16(pkt_size -
- sizeof(*eth_hdr) -
- sizeof(*ip_hdr));
- pkt->nb_segs = 1;
- pkt->pkt_len = pkt_size;
- pkt->ol_flags &= EXT_ATTACHED_MBUF;
- pkt->ol_flags |= ol_flags;
- pkt->vlan_tci = vlan_tci;
- pkt->vlan_tci_outer = vlan_tci_outer;
- pkt->l2_len = sizeof(struct rte_ether_hdr);
- pkt->l3_len = sizeof(struct rte_ipv4_hdr);
- pkts_burst[nb_pkt] = pkt;
+ if (!nb_pkt || !nb_clones) {
+ nb_clones = nb_pkt_flowgen_clones;
+ /* Logic limitation */
+ if (nb_clones > nb_pkt_per_burst)
+ nb_clones = nb_pkt_per_burst;
+
+ pkt = rte_mbuf_raw_alloc(mbp);
+ if (!pkt)
+ break;
+
+ pkt->data_len = pkt_size;
+ pkt->next = NULL;
+
+ /* Initialize Ethernet header. */
+ eth_hdr = rte_pktmbuf_mtod(pkt, struct rte_ether_hdr *);
+ rte_ether_addr_copy(&cfg_ether_dst, ð_hdr->d_addr);
+ rte_ether_addr_copy(&cfg_ether_src, ð_hdr->s_addr);
+ eth_hdr->ether_type = rte_cpu_to_be_16(RTE_ETHER_TYPE_IPV4);
+
+ /* Initialize IP header. */
+ ip_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
+ memset(ip_hdr, 0, sizeof(*ip_hdr));
+ ip_hdr->version_ihl = RTE_IPV4_VHL_DEF;
+ ip_hdr->type_of_service = 0;
+ ip_hdr->fragment_offset = 0;
+ ip_hdr->time_to_live = IP_DEFTTL;
+ ip_hdr->next_proto_id = IPPROTO_UDP;
+ ip_hdr->packet_id = 0;
+ ip_hdr->src_addr = rte_cpu_to_be_32(cfg_ip_src);
+ ip_hdr->dst_addr = rte_cpu_to_be_32(cfg_ip_dst +
+ next_flow);
+ ip_hdr->total_length = RTE_CPU_TO_BE_16(pkt_size -
+ sizeof(*eth_hdr));
+ ip_hdr->hdr_checksum = ip_sum((unaligned_uint16_t *)ip_hdr,
+ sizeof(*ip_hdr));
+
+ /* Initialize UDP header. */
+ udp_hdr = (struct rte_udp_hdr *)(ip_hdr + 1);
+ udp_hdr->src_port = rte_cpu_to_be_16(cfg_udp_src);
+ udp_hdr->dst_port = rte_cpu_to_be_16(cfg_udp_dst);
+ udp_hdr->dgram_cksum = 0; /* No UDP checksum. */
+ udp_hdr->dgram_len = RTE_CPU_TO_BE_16(pkt_size -
+ sizeof(*eth_hdr) -
+ sizeof(*ip_hdr));
+ pkt->nb_segs = 1;
+ pkt->pkt_len = pkt_size;
+ pkt->ol_flags &= EXT_ATTACHED_MBUF;
+ pkt->ol_flags |= ol_flags;
+ pkt->vlan_tci = vlan_tci;
+ pkt->vlan_tci_outer = vlan_tci_outer;
+ pkt->l2_len = sizeof(struct rte_ether_hdr);
+ pkt->l3_len = sizeof(struct rte_ipv4_hdr);
+ } else {
+ nb_clones--;
+ rte_mbuf_refcnt_update(pkt, 1);
+ }
+ pkts_burst[nb_pkt] = pkt;
next_flow = (next_flow + 1) % cfg_n_flows;
}
@@ -163,6 +163,7 @@ usage(char* progname)
printf(" --hairpinq=N: set the number of hairpin queues per port to "
"N.\n");
printf(" --burst=N: set the number of packets per burst to N.\n");
+ printf(" --flowgen-clones=N: set the number of single packet clones to send in flowgen mode. Should be less than burst value.\n");
printf(" --mbcache=N: set the cache of mbuf memory pool to N.\n");
printf(" --rxpt=N: set prefetch threshold register of RX rings to N.\n");
printf(" --rxht=N: set the host threshold register of RX rings to N.\n");
@@ -561,6 +562,7 @@ launch_args_parse(int argc, char** argv)
{ "hairpinq", 1, 0, 0 },
{ "hairpin-mode", 1, 0, 0 },
{ "burst", 1, 0, 0 },
+ { "flowgen-clones", 1, 0, 0 },
{ "mbcache", 1, 0, 0 },
{ "txpt", 1, 0, 0 },
{ "txht", 1, 0, 0 },
@@ -1089,6 +1091,14 @@ launch_args_parse(int argc, char** argv)
else
nb_pkt_per_burst = (uint16_t) n;
}
+ if (!strcmp(lgopts[opt_idx].name, "flowgen-clones")) {
+ n = atoi(optarg);
+ if (n >= 0)
+ nb_pkt_flowgen_clones = (uint16_t) n;
+ else
+ rte_exit(EXIT_FAILURE,
+ "clones must be >= 0 and <= current burst\n");
+ }
if (!strcmp(lgopts[opt_idx].name, "mbcache")) {
n = atoi(optarg);
if ((n >= 0) &&
@@ -240,6 +240,7 @@ uint32_t tx_pkt_times_intra;
/**< Timings for send scheduling in TXONLY mode, time between packets. */
uint16_t nb_pkt_per_burst = DEF_PKT_BURST; /**< Number of packets per burst. */
+uint16_t nb_pkt_flowgen_clones; /**< Number of tx packet clones to send in flowgen mode. */
uint16_t mb_mempool_cache = DEF_MBUF_CACHE; /**< Size of mbuf mempool cache. */
/* current configuration is in DCB or not,0 means it is not in DCB mode */
@@ -476,6 +476,7 @@ extern enum tx_pkt_split tx_pkt_split;
extern uint8_t txonly_multi_flow;
extern uint16_t nb_pkt_per_burst;
+extern uint16_t nb_pkt_flowgen_clones;
extern uint16_t mb_mempool_cache;
extern int8_t rx_pthresh;
extern int8_t rx_hthresh;
@@ -299,6 +299,13 @@ The command line options are:
If set to 0, driver default is used if defined. Else, if driver
default is not defined, default of 32 is used.
+* ``--flowgen-clones=N``
+
+ Set the number of each packet clones to be sent in `flowgen` mode.
+ Sending clones reduces host CPU load on creating packets and may help
+ in testing extreme speeds or maxing out tx packet performance.
+ N should be not zero, but less than 'burst' parameter.
+
* ``--mbcache=N``
Set the cache of mbuf memory pools to N, where 0 <= N <= 512.