3 HOWTO for the linux packet generator
4 ------------------------------------
6 Enable CONFIG_NET_PKTGEN to compile and build pktgen either in-kernel
7 or as a module. A module is preferred; modprobe pktgen if needed. Once
8 running, pktgen creates a thread for each CPU with affinity to that CPU.
9 Monitoring and controlling is done via /proc. It is easiest to select a
10 suitable sample script and configure that.
15 root 129 0.3 0.0 0 0 ? SW 2003 523:20 [pktgen/0]
16 root 130 0.3 0.0 0 0 ? SW 2003 509:50 [pktgen/1]
19 For monitoring and control pktgen creates:
20 /proc/net/pktgen/pgctrl
21 /proc/net/pktgen/kpktgend_X
25 Tuning NIC for max performance
26 ==============================
28 The default NIC settings are (likely) not tuned for pktgen's artificial
29 overload type of benchmarking, as this could hurt the normal use-case.
31 Specifically increasing the TX ring buffer in the NIC:
32 # ethtool -G ethX tx 1024
34 A larger TX ring can improve pktgen's performance, while it can hurt
35 in the general case, 1) because the TX ring buffer might get larger
36 than the CPU's L1/L2 cache, 2) because it allows more queueing in the
37 NIC HW layer (which is bad for bufferbloat).
39 One should hesitate to conclude that packets/descriptors in the HW
40 TX ring cause delay. Drivers usually delay cleaning up the
41 ring-buffers for various performance reasons, and packets stalling
42 the TX ring might just be waiting for cleanup.
44 This cleanup issue is specifically the case for the driver ixgbe
45 (Intel 82599 chip). This driver (ixgbe) combines TX+RX ring cleanups,
46 and the cleanup interval is affected by the ethtool --coalesce setting
47 of parameter "rx-usecs".
49 For ixgbe use e.g. "30" resulting in approx 33K interrupts/sec (1/30*10^6):
50 # ethtool -C ethX rx-usecs 30
55 Pktgen creates a thread for each CPU with affinity to that CPU.
56 Which is controlled through procfile /proc/net/pktgen/kpktgend_X.
58 Example: /proc/net/pktgen/kpktgend_0
62 Result: OK: add_device=eth4@0
64 Most important are the devices assigned to the thread.
66 The two basic thread commands are:
67 * add_device DEVICE@NAME -- adds a single device
68 * rem_device_all -- remove all associated devices
70 When adding a device to a thread, a corresponding procfile is created
71 which is used for configuring this device. Thus, device names need to
74 To support adding the same device to multiple threads, which is useful
75 with multi queue NICs, the device naming scheme is extended with "@":
78 The part after "@" can be anything, but it is custom to use the thread
84 The Params section holds configured information. The Current section
85 holds running statistics. The Result is printed after a run or after
86 interruption. Example:
88 /proc/net/pktgen/eth4@0
90 Params: count 100000 min_pkt_size: 60 max_pkt_size: 60
91 frags: 0 delay: 0 clone_skb: 64 ifname: eth4@0
93 queue_map_min: 0 queue_map_max: 0
94 dst_min: 192.168.81.2 dst_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
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
108 Result: OK: 15430(c15405+d25) usec, 100000 (60byte,0frags)
109 6480562pps 3110Mb/sec (3110669760bps) errors: 0
114 This is done via the /proc interface, and most easily done via pgset
115 as defined in the sample scripts.
119 pgset "clone_skb 1" sets the number of copies of the same packet
120 pgset "clone_skb 0" use single SKB for all transmits
121 pgset "burst 8" uses xmit_more API to queue 8 copies of the same
122 packet and update HW tx queue tail pointer once.
123 "burst 1" is the default
124 pgset "pkt_size 9014" sets packet size to 9014
125 pgset "frags 5" packet will consist of 5 fragments
126 pgset "count 200000" sets number of packets to send, set to zero
127 for continuous sends until explicitly stopped.
129 pgset "delay 5000" adds delay to hard_start_xmit(). nanoseconds
131 pgset "dst 10.0.0.1" sets IP destination address
132 (BEWARE! This generator is very aggressive!)
134 pgset "dst_min 10.0.0.1" Same as dst
135 pgset "dst_max 10.0.0.254" Set the maximum destination IP.
136 pgset "src_min 10.0.0.1" Set the minimum (or only) source IP.
137 pgset "src_max 10.0.0.254" Set the maximum source IP.
138 pgset "dst6 fec0::1" IPV6 destination address
139 pgset "src6 fec0::2" IPV6 source address
140 pgset "dstmac 00:00:00:00:00:00" sets MAC destination address
141 pgset "srcmac 00:00:00:00:00:00" sets MAC source address
143 pgset "queue_map_min 0" Sets the min value of tx queue interval
144 pgset "queue_map_max 7" Sets the max value of tx queue interval, for multiqueue devices
145 To select queue 1 of a given device,
146 use queue_map_min=1 and queue_map_max=1
148 pgset "src_mac_count 1" Sets the number of MACs we'll range through.
149 The 'minimum' MAC is what you set with srcmac.
151 pgset "dst_mac_count 1" Sets the number of MACs we'll range through.
152 The 'minimum' MAC is what you set with dstmac.
154 pgset "flag [name]" Set a flag to determine behaviour. Current flags
155 are: IPSRC_RND # IP source is random (between min/max)
156 IPDST_RND # IP destination is random
157 UDPSRC_RND, UDPDST_RND,
158 MACSRC_RND, MACDST_RND
160 MPLS_RND, VID_RND, SVID_RND
162 QUEUE_MAP_RND # queue map random
163 QUEUE_MAP_CPU # queue map mirrors smp_processor_id()
165 IPSEC # IPsec encapsulation (needs CONFIG_XFRM)
166 NODE_ALLOC # node specific memory allocation
167 NO_TIMESTAMP # disable timestamping
169 pgset spi SPI_VALUE Set specific SA used to transform packet.
171 pgset "udp_src_min 9" set UDP source port min, If < udp_src_max, then
172 cycle through the port range.
174 pgset "udp_src_max 9" set UDP source port max.
175 pgset "udp_dst_min 9" set UDP destination port min, If < udp_dst_max, then
176 cycle through the port range.
177 pgset "udp_dst_max 9" set UDP destination port max.
179 pgset "mpls 0001000a,0002000a,0000000a" set MPLS labels (in this example
180 outer label=16,middle label=32,
181 inner label=0 (IPv4 NULL)) Note that
182 there must be no spaces between the
183 arguments. Leading zeros are required.
184 Do not set the bottom of stack bit,
185 that's done automatically. If you do
186 set the bottom of stack bit, that
187 indicates that you want to randomly
188 generate that address and the flag
189 MPLS_RND will be turned on. You
190 can have any mix of random and fixed
191 labels in the label stack.
193 pgset "mpls 0" turn off mpls (or any invalid argument works too!)
195 pgset "vlan_id 77" set VLAN ID 0-4095
196 pgset "vlan_p 3" set priority bit 0-7 (default 0)
197 pgset "vlan_cfi 0" set canonical format identifier 0-1 (default 0)
199 pgset "svlan_id 22" set SVLAN ID 0-4095
200 pgset "svlan_p 3" set priority bit 0-7 (default 0)
201 pgset "svlan_cfi 0" set canonical format identifier 0-1 (default 0)
203 pgset "vlan_id 9999" > 4095 remove vlan and svlan tags
204 pgset "svlan 9999" > 4095 remove svlan tag
207 pgset "tos XX" set former IPv4 TOS field (e.g. "tos 28" for AF11 no ECN, default 00)
208 pgset "traffic_class XX" set former IPv6 TRAFFIC CLASS (e.g. "traffic_class B8" for EF no ECN, default 00)
210 pgset stop aborts injection. Also, ^C aborts generator.
212 pgset "rate 300M" set rate to 300 Mb/s
213 pgset "ratep 1000000" set rate to 1Mpps
215 pgset "xmit_mode netif_receive" RX inject into stack netif_receive_skb()
216 Works with "burst" but not with "clone_skb".
217 Default xmit_mode is "start_xmit".
222 A collection of tutorial scripts and helpers for pktgen is in the
223 samples/pktgen directory. The helper parameters.sh file support easy
224 and consistent parameter parsing across the sample scripts.
226 Usage example and help:
227 ./pktgen_sample01_simple.sh -i eth4 -m 00:1B:21:3C:9D:F8 -d 192.168.8.2
229 Usage: ./pktgen_sample01_simple.sh [-vx] -i ethX
230 -i : ($DEV) output interface/device (required)
231 -s : ($PKT_SIZE) packet size
232 -d : ($DEST_IP) destination IP
233 -m : ($DST_MAC) destination MAC-addr
234 -t : ($THREADS) threads to start
235 -c : ($SKB_CLONE) SKB clones send before alloc new SKB
236 -b : ($BURST) HW level bursting of SKBs
237 -v : ($VERBOSE) verbose
240 The global variables being set are also listed. E.g. the required
241 interface/device parameter "-i" sets variable $DEV. Copy the
242 pktgen_sampleXX scripts and modify them to fit your own needs.
246 pktgen.conf-1-2 # 1 CPU 2 dev
247 pktgen.conf-1-1-rdos # 1 CPU 1 dev w. route DoS
248 pktgen.conf-1-1-ip6 # 1 CPU 1 dev ipv6
249 pktgen.conf-1-1-ip6-rdos # 1 CPU 1 dev ipv6 w. route DoS
250 pktgen.conf-1-1-flows # 1 CPU 1 dev multiple flows.
255 Note that when adding devices to a specific CPU it is a good idea to
256 also assign /proc/irq/XX/smp_affinity so that the TX interrupts are bound
257 to the same CPU. This reduces cache bouncing when freeing skbs.
259 Plus using the device flag QUEUE_MAP_CPU, which maps the SKBs TX queue
260 to the running threads CPU (directly from smp_processor_id()).
264 Default IPsec transformation with ESP encapsulation plus transport mode
265 can be enabled by simply setting:
270 To avoid breaking existing testbed scripts for using AH type and tunnel mode,
271 you can use "pgset spi SPI_VALUE" to specify which transformation mode
275 Current commands and configuration options
276 ==========================================
278 ** Pgcontrol commands:
368 xmit_mode <start_xmit|netif_receive>
380 ftp://robur.slu.se/pub/Linux/net-development/pktgen-testing/
381 ftp://robur.slu.se/pub/Linux/net-development/pktgen-testing/examples/
383 Paper from Linux-Kongress in Erlangen 2004.
384 ftp://robur.slu.se/pub/Linux/net-development/pktgen-testing/pktgen_paper.pdf
387 Grant Grundler for testing on IA-64 and parisc, Harald Welte, Lennert Buytenhek
388 Stephen Hemminger, Andi Kleen, Dave Miller and many others.
391 Good luck with the linux net-development.