1 2 3 HOWTO for the linux packet generator 4 ------------------------------------ 5 6Enable CONFIG_NET_PKTGEN to compile and build pktgen either in-kernel 7or as a module. A module is preferred; modprobe pktgen if needed. Once 8running, pktgen creates a thread for each CPU with affinity to that CPU. 9Monitoring and controlling is done via /proc. It is easiest to select a 10suitable sample script and configure that. 11 12On a dual CPU: 13 14ps aux | grep pkt 15root 129 0.3 0.0 0 0 ? SW 2003 523:20 [kpktgend_0] 16root 130 0.3 0.0 0 0 ? SW 2003 509:50 [kpktgend_1] 17 18 19For monitoring and control pktgen creates: 20 /proc/net/pktgen/pgctrl 21 /proc/net/pktgen/kpktgend_X 22 /proc/net/pktgen/ethX 23 24 25Tuning NIC for max performance 26============================== 27 28The default NIC settings are (likely) not tuned for pktgen's artificial 29overload type of benchmarking, as this could hurt the normal use-case. 30 31Specifically increasing the TX ring buffer in the NIC: 32 # ethtool -G ethX tx 1024 33 34A larger TX ring can improve pktgen's performance, while it can hurt 35in the general case, 1) because the TX ring buffer might get larger 36than the CPU's L1/L2 cache, 2) because it allows more queueing in the 37NIC HW layer (which is bad for bufferbloat). 38 39One should hesitate to conclude that packets/descriptors in the HW 40TX ring cause delay. Drivers usually delay cleaning up the 41ring-buffers for various performance reasons, and packets stalling 42the TX ring might just be waiting for cleanup. 43 44This cleanup issue is specifically the case for the driver ixgbe 45(Intel 82599 chip). This driver (ixgbe) combines TX+RX ring cleanups, 46and the cleanup interval is affected by the ethtool --coalesce setting 47of parameter "rx-usecs". 48 49For ixgbe use e.g. "30" resulting in approx 33K interrupts/sec (1/30*10^6): 50 # ethtool -C ethX rx-usecs 30 51 52 53Kernel threads 54============== 55Pktgen creates a thread for each CPU with affinity to that CPU. 56Which is controlled through procfile /proc/net/pktgen/kpktgend_X. 57 58Example: /proc/net/pktgen/kpktgend_0 59 60 Running: 61 Stopped: eth4@0 62 Result: OK: add_device=eth4@0 63 64Most important are the devices assigned to the thread. 65 66The two basic thread commands are: 67 * add_device DEVICE@NAME -- adds a single device 68 * rem_device_all -- remove all associated devices 69 70When adding a device to a thread, a corresponding procfile is created 71which is used for configuring this device. Thus, device names need to 72be unique. 73 74To support adding the same device to multiple threads, which is useful 75with multi queue NICs, the device naming scheme is extended with "@": 76 device@something 77 78The part after "@" can be anything, but it is custom to use the thread 79number. 80 81Viewing devices 82=============== 83 84The Params section holds configured information. The Current section 85holds running statistics. The Result is printed after a run or after 86interruption. Example: 87 88/proc/net/pktgen/eth4@0 89 90 Params: count 100000 min_pkt_size: 60 max_pkt_size: 60 91 frags: 0 delay: 0 clone_skb: 64 ifname: eth4@0 92 flows: 0 flowlen: 0 93 queue_map_min: 0 queue_map_max: 0 94 dst_min: 192.168.81.2 dst_max: 95 src_min: src_max: 96 src_mac: 90:e2:ba:0a:56:b4 dst_mac: 00:1b:21:3c:9d:f8 97 udp_src_min: 9 udp_src_max: 109 udp_dst_min: 9 udp_dst_max: 9 98 src_mac_count: 0 dst_mac_count: 0 99 Flags: UDPSRC_RND NO_TIMESTAMP QUEUE_MAP_CPU 100 Current: 101 pkts-sofar: 100000 errors: 0 102 started: 623913381008us stopped: 623913396439us idle: 25us 103 seq_num: 100001 cur_dst_mac_offset: 0 cur_src_mac_offset: 0 104 cur_saddr: 192.168.8.3 cur_daddr: 192.168.81.2 105 cur_udp_dst: 9 cur_udp_src: 42 106 cur_queue_map: 0 107 flows: 0 108 Result: OK: 15430(c15405+d25) usec, 100000 (60byte,0frags) 109 6480562pps 3110Mb/sec (3110669760bps) errors: 0 110 111 112Configuring devices 113=================== 114This is done via the /proc interface, and most easily done via pgset 115as defined in the sample scripts. 116You need to specify PGDEV environment variable to use functions from sample 117scripts, i.e.: 118export PGDEV=/proc/net/pktgen/eth4@0 119source samples/pktgen/functions.sh 120 121Examples: 122 123 pg_ctrl start starts injection. 124 pg_ctrl stop aborts injection. Also, ^C aborts generator. 125 126 pgset "clone_skb 1" sets the number of copies of the same packet 127 pgset "clone_skb 0" use single SKB for all transmits 128 pgset "burst 8" uses xmit_more API to queue 8 copies of the same 129 packet and update HW tx queue tail pointer once. 130 "burst 1" is the default 131 pgset "pkt_size 9014" sets packet size to 9014 132 pgset "frags 5" packet will consist of 5 fragments 133 pgset "count 200000" sets number of packets to send, set to zero 134 for continuous sends until explicitly stopped. 135 136 pgset "delay 5000" adds delay to hard_start_xmit(). nanoseconds 137 138 pgset "dst 10.0.0.1" sets IP destination address 139 (BEWARE! This generator is very aggressive!) 140 141 pgset "dst_min 10.0.0.1" Same as dst 142 pgset "dst_max 10.0.0.254" Set the maximum destination IP. 143 pgset "src_min 10.0.0.1" Set the minimum (or only) source IP. 144 pgset "src_max 10.0.0.254" Set the maximum source IP. 145 pgset "dst6 fec0::1" IPV6 destination address 146 pgset "src6 fec0::2" IPV6 source address 147 pgset "dstmac 00:00:00:00:00:00" sets MAC destination address 148 pgset "srcmac 00:00:00:00:00:00" sets MAC source address 149 150 pgset "queue_map_min 0" Sets the min value of tx queue interval 151 pgset "queue_map_max 7" Sets the max value of tx queue interval, for multiqueue devices 152 To select queue 1 of a given device, 153 use queue_map_min=1 and queue_map_max=1 154 155 pgset "src_mac_count 1" Sets the number of MACs we'll range through. 156 The 'minimum' MAC is what you set with srcmac. 157 158 pgset "dst_mac_count 1" Sets the number of MACs we'll range through. 159 The 'minimum' MAC is what you set with dstmac. 160 161 pgset "flag [name]" Set a flag to determine behaviour. Current flags 162 are: IPSRC_RND # IP source is random (between min/max) 163 IPDST_RND # IP destination is random 164 UDPSRC_RND, UDPDST_RND, 165 MACSRC_RND, MACDST_RND 166 TXSIZE_RND, IPV6, 167 MPLS_RND, VID_RND, SVID_RND 168 FLOW_SEQ, 169 QUEUE_MAP_RND # queue map random 170 QUEUE_MAP_CPU # queue map mirrors smp_processor_id() 171 UDPCSUM, 172 IPSEC # IPsec encapsulation (needs CONFIG_XFRM) 173 NODE_ALLOC # node specific memory allocation 174 NO_TIMESTAMP # disable timestamping 175 pgset 'flag ![name]' Clear a flag to determine behaviour. 176 Note that you might need to use single quote in 177 interactive mode, so that your shell wouldn't expand 178 the specified flag as a history command. 179 180 pgset "spi [SPI_VALUE]" Set specific SA used to transform packet. 181 182 pgset "udp_src_min 9" set UDP source port min, If < udp_src_max, then 183 cycle through the port range. 184 185 pgset "udp_src_max 9" set UDP source port max. 186 pgset "udp_dst_min 9" set UDP destination port min, If < udp_dst_max, then 187 cycle through the port range. 188 pgset "udp_dst_max 9" set UDP destination port max. 189 190 pgset "mpls 0001000a,0002000a,0000000a" set MPLS labels (in this example 191 outer label=16,middle label=32, 192 inner label=0 (IPv4 NULL)) Note that 193 there must be no spaces between the 194 arguments. Leading zeros are required. 195 Do not set the bottom of stack bit, 196 that's done automatically. If you do 197 set the bottom of stack bit, that 198 indicates that you want to randomly 199 generate that address and the flag 200 MPLS_RND will be turned on. You 201 can have any mix of random and fixed 202 labels in the label stack. 203 204 pgset "mpls 0" turn off mpls (or any invalid argument works too!) 205 206 pgset "vlan_id 77" set VLAN ID 0-4095 207 pgset "vlan_p 3" set priority bit 0-7 (default 0) 208 pgset "vlan_cfi 0" set canonical format identifier 0-1 (default 0) 209 210 pgset "svlan_id 22" set SVLAN ID 0-4095 211 pgset "svlan_p 3" set priority bit 0-7 (default 0) 212 pgset "svlan_cfi 0" set canonical format identifier 0-1 (default 0) 213 214 pgset "vlan_id 9999" > 4095 remove vlan and svlan tags 215 pgset "svlan 9999" > 4095 remove svlan tag 216 217 218 pgset "tos XX" set former IPv4 TOS field (e.g. "tos 28" for AF11 no ECN, default 00) 219 pgset "traffic_class XX" set former IPv6 TRAFFIC CLASS (e.g. "traffic_class B8" for EF no ECN, default 00) 220 221 pgset "rate 300M" set rate to 300 Mb/s 222 pgset "ratep 1000000" set rate to 1Mpps 223 224 pgset "xmit_mode netif_receive" RX inject into stack netif_receive_skb() 225 Works with "burst" but not with "clone_skb". 226 Default xmit_mode is "start_xmit". 227 228Sample scripts 229============== 230 231A collection of tutorial scripts and helpers for pktgen is in the 232samples/pktgen directory. The helper parameters.sh file support easy 233and consistent parameter parsing across the sample scripts. 234 235Usage example and help: 236 ./pktgen_sample01_simple.sh -i eth4 -m 00:1B:21:3C:9D:F8 -d 192.168.8.2 237 238Usage: ./pktgen_sample01_simple.sh [-vx] -i ethX 239 -i : ($DEV) output interface/device (required) 240 -s : ($PKT_SIZE) packet size 241 -d : ($DEST_IP) destination IP 242 -m : ($DST_MAC) destination MAC-addr 243 -t : ($THREADS) threads to start 244 -c : ($SKB_CLONE) SKB clones send before alloc new SKB 245 -b : ($BURST) HW level bursting of SKBs 246 -v : ($VERBOSE) verbose 247 -x : ($DEBUG) debug 248 249The global variables being set are also listed. E.g. the required 250interface/device parameter "-i" sets variable $DEV. Copy the 251pktgen_sampleXX scripts and modify them to fit your own needs. 252 253The old scripts: 254 255pktgen.conf-1-2 # 1 CPU 2 dev 256pktgen.conf-1-1-rdos # 1 CPU 1 dev w. route DoS 257pktgen.conf-1-1-ip6 # 1 CPU 1 dev ipv6 258pktgen.conf-1-1-ip6-rdos # 1 CPU 1 dev ipv6 w. route DoS 259pktgen.conf-1-1-flows # 1 CPU 1 dev multiple flows. 260 261 262Interrupt affinity 263=================== 264Note that when adding devices to a specific CPU it is a good idea to 265also assign /proc/irq/XX/smp_affinity so that the TX interrupts are bound 266to the same CPU. This reduces cache bouncing when freeing skbs. 267 268Plus using the device flag QUEUE_MAP_CPU, which maps the SKBs TX queue 269to the running threads CPU (directly from smp_processor_id()). 270 271Enable IPsec 272============ 273Default IPsec transformation with ESP encapsulation plus transport mode 274can be enabled by simply setting: 275 276pgset "flag IPSEC" 277pgset "flows 1" 278 279To avoid breaking existing testbed scripts for using AH type and tunnel mode, 280you can use "pgset spi SPI_VALUE" to specify which transformation mode 281to employ. 282 283 284Current commands and configuration options 285========================================== 286 287** Pgcontrol commands: 288 289start 290stop 291reset 292 293** Thread commands: 294 295add_device 296rem_device_all 297 298 299** Device commands: 300 301count 302clone_skb 303burst 304debug 305 306frags 307delay 308 309src_mac_count 310dst_mac_count 311 312pkt_size 313min_pkt_size 314max_pkt_size 315 316queue_map_min 317queue_map_max 318skb_priority 319 320tos (ipv4) 321traffic_class (ipv6) 322 323mpls 324 325udp_src_min 326udp_src_max 327 328udp_dst_min 329udp_dst_max 330 331node 332 333flag 334 IPSRC_RND 335 IPDST_RND 336 UDPSRC_RND 337 UDPDST_RND 338 MACSRC_RND 339 MACDST_RND 340 TXSIZE_RND 341 IPV6 342 MPLS_RND 343 VID_RND 344 SVID_RND 345 FLOW_SEQ 346 QUEUE_MAP_RND 347 QUEUE_MAP_CPU 348 UDPCSUM 349 IPSEC 350 NODE_ALLOC 351 NO_TIMESTAMP 352 353spi (ipsec) 354 355dst_min 356dst_max 357 358src_min 359src_max 360 361dst_mac 362src_mac 363 364clear_counters 365 366src6 367dst6 368dst6_max 369dst6_min 370 371flows 372flowlen 373 374rate 375ratep 376 377xmit_mode <start_xmit|netif_receive> 378 379vlan_cfi 380vlan_id 381vlan_p 382 383svlan_cfi 384svlan_id 385svlan_p 386 387 388References: 389ftp://robur.slu.se/pub/Linux/net-development/pktgen-testing/ 390ftp://robur.slu.se/pub/Linux/net-development/pktgen-testing/examples/ 391 392Paper from Linux-Kongress in Erlangen 2004. 393ftp://robur.slu.se/pub/Linux/net-development/pktgen-testing/pktgen_paper.pdf 394 395Thanks to: 396Grant Grundler for testing on IA-64 and parisc, Harald Welte, Lennert Buytenhek 397Stephen Hemminger, Andi Kleen, Dave Miller and many others. 398 399 400Good luck with the linux net-development. 401