aboutsummaryrefslogtreecommitdiffstats
path: root/docs/report/vpp_performance_tests/overview.rst
blob: a28d8572eb2f8997b3840b2c674a3619618f20fa (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
Overview
========

Tested Physical Topologies
--------------------------

CSIT VPP performance tests are executed on physical baremetal servers hosted by
LF FD.io project. Testbed physical topology is shown in the figure below.

::

    +------------------------+           +------------------------+
    |                        |           |                        |
    |  +------------------+  |           |  +------------------+  |
    |  |                  |  |           |  |                  |  |
    |  |                  <----------------->                  |  |
    |  |       DUT1       |  |           |  |       DUT2       |  |
    |  +--^---------------+  |           |  +---------------^--+  |
    |     |                  |           |                  |     |
    |     |            SUT1  |           |  SUT2            |     |
    +------------------------+           +------------------^-----+
          |                                                 |
          |                                                 |
          |                  +-----------+                  |
          |                  |           |                  |
          +------------------>    TG     <------------------+
                             |           |
                             +-----------+

SUT1 and SUT2 are two System Under Test servers (Cisco UCS C240, each with two
Intel XEON CPUs), TG is a Traffic Generator (TG, another Cisco UCS C240, with
two Intel XEON CPUs). SUTs run VPP SW application in Linux user-mode as a
Device Under Test (DUT). TG runs TRex SW application as a packet Traffic
Generator. Physical connectivity between SUTs and to TG is provided using
different NIC models that need to be tested for performance. Currently
installed and tested NIC models include:

#. 2port10GE X520-DA2 Intel.
#. 2port10GE X710 Intel.
#. 2port10GE VIC1227 Cisco.
#. 2port40GE VIC1385 Cisco.
#. 2port40GE XL710 Intel.

From SUT and DUT perspective, all performance tests involve forwarding packets
between two physical Ethernet ports (10GE or 40GE). Due to the number of
listed NIC models tested and available PCI slot capacity in SUT servers, in
all of the above cases both physical ports are located on the same NIC. In
some test cases this results in measured packet throughput being limited not
by VPP DUT but by either the physical interface or the NIC capacity.

Going forward CSIT project will be looking to add more hardware into FD.io
performance labs to address larger scale multi-interface and multi-NIC
performance testing scenarios.

For test cases that require DUT (VPP) to communicate with VM(s) over vhost-user
interfaces, N of VM instances are created on SUT1 and SUT2. For N=1 DUT (VPP)
forwards packets between vhostuser and physical interfaces. For N>1 DUT (VPP) a
logical service chain forwarding topology is created on DUT (VPP) by applying L2
or IPv4/IPv6 configuration depending on the test suite.
DUT (VPP) test topology with N VM instances
is shown in the figure below including applicable packet flow thru the DUTs and
VMs (marked in the figure with ``***``).

::

    +-------------------------+           +-------------------------+
    | +---------+ +---------+ |           | +---------+ +---------+ |
    | |  VM[1]  | |  VM[N]  | |           | |  VM[1]  | |  VM[N]  | |
    | |  *****  | |  *****  | |           | |  *****  | |  *****  | |
    | +--^---^--+ +--^---^--+ |           | +--^---^--+ +--^---^--+ |
    |   *|   |*     *|   |*   |           |   *|   |*     *|   |*   |
    | +--v---v-------v---v--+ |           | +--v---v-------v---v--+ |
    | |  *   *       *   *  |*|***********|*|  *   *       *   *  | |
    | |  *   *********   ***<-|-----------|->***   *********   *  | |
    | |  *    DUT1          | |           | |       DUT2       *  | |
    | +--^------------------+ |           | +------------------^--+ |
    |   *|                    |           |                    |*   |
    |   *|            SUT1    |           |  SUT2              |*   |
    +-------------------------+           +-------------------------+
        *|                                                     |*
        *|                                                     |*
        *|                    +-----------+                    |*
        *|                    |           |                    |*
        *+-------------------->    TG     <--------------------+*
        **********************|           |**********************
                              +-----------+

For VM tests, packets are switched by DUT (VPP) multiple times: twice for a
single VM, three times for two VMs, N+1 times for N VMs.
Hence the external
throughput rates measured by TG and listed in this report must be multiplied
by (N+1) to represent the actual DUT aggregate packet forwarding rate.

Note that reported VPP performance results are specific to the SUTs tested.
Current LF FD.io SUTs are based on Intel XEON E5-2699v3 2.3GHz CPUs. SUTs with
other CPUs are likely to yield different results. A good rule of thumb, that
can be applied to estimate VPP packet thoughput for Phy-to-Phy (NIC-to-NIC,
PCI-to-PCI) topology, is to expect the forwarding performance to be
proportional to CPU core frequency, assuming CPU is the only limiting factor
and all other SUT parameters equivalent to FD.io CSIT environment. The same rule
of thumb can be also applied for Phy-to-VM-to-Phy (NIC-to-VM-to-NIC) topology,
but due to much higher dependency on intensive memory operations and
sensitivity to Linux kernel scheduler settings and behaviour, this estimation
may not always yield good enough accuracy.

For detailed LF FD.io test bed specification and physical topology please refer
to `LF FDio CSIT testbed wiki page
<https://wiki.fd.io/view/CSIT/CSIT_LF_testbed>`_.

Performance Tests Coverage
--------------------------

Performance tests are split into the two main categories:

- Throughput discovery - discovery of packet forwarding rate using binary search
  in accordance to RFC2544.

  - NDR - discovery of Non Drop Rate packet throughput, at zero packet loss;
    followed by one-way packet latency measurements at 10%, 50% and 100% of
    discovered NDR throughput.
  - PDR - discovery of Partial Drop Rate, with specified non-zero packet loss
    currently set to 0.5%; followed by one-way packet latency measurements at
    100% of discovered PDR throughput.

- Throughput verification - verification of packet forwarding rate against
  previously discovered throughput rate. These tests are currently done against
  0.9 of reference NDR, with reference rates updated periodically.

CSIT |release| includes following performance test suites, listed per NIC type:

- 2port10GE X520-DA2 Intel

  - **L2XC** - L2 Cross-Connect switched-forwarding of untagged, dot1q, dot1ad
    VLAN tagged Ethernet frames.
  - **L2BD** - L2 Bridge-Domain switched-forwarding of untagged Ethernet frames
    with MAC learning; disabled MAC learning i.e. static MAC tests to be added.
  - **IPv4** - IPv4 routed-forwarding.
  - **IPv6** - IPv6 routed-forwarding.
  - **IPv4 Scale** - IPv4 routed-forwarding with 20k, 200k and 2M FIB entries.
  - **IPv6 Scale** - IPv6 routed-forwarding with 20k, 200k and 2M FIB entries.
  - **VMs with vhost-user** - virtual topologies with 1 VM and service chains
    of 2 VMs using vhost-user interfaces, with VPP forwarding modes incl. L2
    Cross-Connect, L2 Bridge-Domain, VXLAN with L2BD, IPv4 routed-forwarding.
  - **COP** - IPv4 and IPv6 routed-forwarding with COP address security.
  - **iACL** - IPv4 and IPv6 routed-forwarding with iACL address security.
  - **LISP** - LISP overlay tunneling for IPv4-over-IPv4, IPv6-over-IPv4,
    IPv6-over-IPv6, IPv4-over-IPv6 in IPv4 and IPv6 routed-forwarding modes.
  - **VXLAN** - VXLAN overlay tunnelling integration with L2XC and L2BD.
  - **QoS Policer** - ingress packet rate measuring, marking and limiting
    (IPv4).
  - **CGNAT** - Carrier Grade Network Address Translation tests with varying
    number of users and ports per user.

- 2port40GE XL710 Intel

  - **L2XC** - L2 Cross-Connect switched-forwarding of untagged Ethernet frames.
  - **L2BD** - L2 Bridge-Domain switched-forwarding of untagged Ethernet frames
    with MAC learning.
  - **IPv4** - IPv4 routed-forwarding.
  - **IPv6** - IPv6 routed-forwarding.
  - **VMs with vhost-user** - virtual topologies with 1 VM and service chains
    of 2 VMs using vhost-user interfaces, with VPP forwarding modes incl. L2
    Cross-Connect, L2 Bridge-Domain, VXLAN with L2BD, IPv4 routed-forwarding.
  - **IPSec** - IPSec encryption with AES-GCM, CBC-SHA1 ciphers, in combination
    with IPv4 routed-forwarding.
  - **IPSec+LISP** - IPSec encryption with CBC-SHA1 ciphers, in combination
    with LISP-GPE overlay tunneling for IPv4-over-IPv4.

- 2port10GE X710 Intel

  - **L2BD** - L2 Bridge-Domain switched-forwarding of untagged Ethernet frames
    with MAC learning.
  - **VMs with vhost-user** - virtual topologies with 1 VM using vhost-user
    interfaces, with VPP forwarding modes incl. L2 Bridge-Domain.

- 2port10GE VIC1227 Cisco

  - **L2BD** - L2 Bridge-Domain switched-forwarding of untagged Ethernet frames
    with MAC learning.

- 2port40GE VIC1385 Cisco

  - **L2BD** - L2 Bridge-Domain switched-forwarding of untagged Ethernet frames
     with MAC learning.

Execution of performance tests takes time, especially the throughput discovery
tests. Due to limited HW testbed resources available within FD.io labs hosted
by Linux Foundation, the number of tests for NICs other than X520 (a.k.a.
Niantic) has been limited to few baseline tests. Over time we expect the HW
testbed resources to grow, and will be adding complete set of performance
tests for all models of hardware to be executed regularly and(or)
continuously.

Performance Tests Naming
------------------------

CSIT |release| follows a common structured naming convention for all
performance and system functional tests, introduced in CSIT |release-1|.

The naming should be intuitive for majority of the tests. Complete
description of CSIT test naming convention is provided on `CSIT test naming wiki
<https://wiki.fd.io/view/CSIT/csit-test-naming>`_.

Here few illustrative examples of the new naming usage for performance test
suites:

#. **Physical port to physical port - a.k.a. NIC-to-NIC, Phy-to-Phy, P2P**

   - *PortNICConfig-WireEncapsulation-PacketForwardingFunction-
     PacketProcessingFunction1-...-PacketProcessingFunctionN-TestType*
   - *10ge2p1x520-dot1q-l2bdbasemaclrn-ndrdisc.robot* => 2 ports of 10GE on
     Intel x520 NIC, dot1q tagged Ethernet, L2 bridge-domain baseline switching
     with MAC learning, NDR throughput discovery.
   - *10ge2p1x520-ethip4vxlan-l2bdbasemaclrn-ndrchk.robot* => 2 ports of 10GE
     on Intel x520 NIC, IPv4 VXLAN Ethernet, L2 bridge-domain baseline
     switching with MAC learning, NDR throughput discovery.
   - *10ge2p1x520-ethip4-ip4base-ndrdisc.robot* => 2 ports of 10GE on Intel
     x520 NIC, IPv4 baseline routed forwarding, NDR throughput discovery.
   - *10ge2p1x520-ethip6-ip6scale200k-ndrdisc.robot* => 2 ports of 10GE on
     Intel x520 NIC, IPv6 scaled up routed forwarding, NDR throughput
     discovery.

#. **Physical port to VM (or VM chain) to physical port - a.k.a. NIC2VM2NIC,
   P2V2P, NIC2VMchain2NIC, P2V2V2P**

   - *PortNICConfig-WireEncapsulation-PacketForwardingFunction-
     PacketProcessingFunction1-...-PacketProcessingFunctionN-VirtEncapsulation-
     VirtPortConfig-VMconfig-TestType*
   - *10ge2p1x520-dot1q-l2bdbasemaclrn-eth-2vhost-1vm-ndrdisc.robot* => 2 ports
     of 10GE on Intel x520 NIC, dot1q tagged Ethernet, L2 bridge-domain
     switching to/from two vhost interfaces and one VM, NDR throughput
     discovery.
   - *10ge2p1x520-ethip4vxlan-l2bdbasemaclrn-eth-2vhost-1vm-ndrdisc.robot* => 2
     ports of 10GE on Intel x520 NIC, IPv4 VXLAN Ethernet, L2 bridge-domain
     switching to/from two vhost interfaces and one VM, NDR throughput
     discovery.
   - *10ge2p1x520-ethip4vxlan-l2bdbasemaclrn-eth-4vhost-2vm-ndrdisc.robot* => 2
     ports of 10GE on Intel x520 NIC, IPv4 VXLAN Ethernet, L2 bridge-domain
     switching to/from four vhost interfaces and two VMs, NDR throughput
     discovery.

Methodology: Multi-Thread and Multi-Core
----------------------------------------

**HyperThreading** - CSIT |release| performance tests are executed with SUT
servers' Intel XEON CPUs configured in HyperThreading Disabled mode (BIOS
settings). This is the simplest configuration used to establish baseline
single-thread single-core SW packet processing and forwarding performance.
Subsequent releases of CSIT will add performance tests with Intel
HyperThreading Enabled (requires BIOS settings change and hard reboot).

**Multi-core Test** - CSIT |release| multi-core tests are executed in the
following VPP thread and core configurations:

#. 1t1c - 1 VPP worker thread on 1 CPU physical core.
#. 2t2c - 2 VPP worker threads on 2 CPU physical cores.

Note that in quite a few test cases running VPP on 2 physical cores hits
the tested NIC I/O bandwidth or packets-per-second limit.

Methodology: Packet Throughput
------------------------------

Following values are measured and reported for packet throughput tests:

- NDR binary search per RFC2544:

  - Packet rate: "RATE: <aggregate packet rate in packets-per-second> pps
    (2x <per direction packets-per-second>)"
  - Aggregate bandwidth: "BANDWIDTH: <aggregate bandwidth in Gigabits per
    second> Gbps (untagged)"

- PDR binary search per RFC2544:

  - Packet rate: "RATE: <aggregate packet rate in packets-per-second> pps (2x
    <per direction packets-per-second>)"
  - Aggregate bandwidth: "BANDWIDTH: <aggregate bandwidth in Gigabits per
    second> Gbps (untagged)"
  - Packet loss tolerance: "LOSS_ACCEPTANCE <accepted percentage of packets
    lost at PDR rate>""

- NDR and PDR are measured for the following L2 frame sizes:

  - IPv4: 64B, IMIX_v4_1 (28x64B,16x570B,4x1518B), 1518B, 9000B.
  - IPv6: 78B, 1518B, 9000B.


Methodology: Packet Latency
---------------------------

TRex Traffic Generator (TG) is used for measuring latency of VPP DUTs. Reported
latency values are measured using following methodology:

- Latency tests are performed at 10%, 50% of discovered NDR rate (non drop rate)
  for each NDR throughput test and packet size (except IMIX).
- TG sends dedicated latency streams, one per direction, each at the rate of
  10kpps at the prescribed packet size; these are sent in addition to the main
  load streams.
- TG reports min/avg/max latency values per stream direction, hence two sets
  of latency values are reported per test case; future release of TRex is
  expected to report latency percentiles.
- Reported latency values are aggregate across two SUTs due to three node
  topology used for all performance tests; for per SUT latency, reported value
  should be divided by two.
- 1usec is the measurement accuracy advertised by TRex TG for the setup used in
  FD.io labs used by CSIT project.
- TRex setup introduces an always-on error of about 2*2usec per latency flow -
  additonal Tx/Rx interface latency induced by TRex SW writing and reading
  packet timestamps on CPU cores without HW acceleration on NICs closer to the
  interface line.


Methodology: KVM VM vhost
-------------------------

CSIT |release| introduced environment configuration changes to KVM Qemu vhost-
user tests in order to more representatively measure |vpp-release| performance
in configurations with vhost-user interfaces and VMs.

Current setup of CSIT FD.io performance lab is using tuned settings for more
optimal performance of KVM Qemu:

- Qemu virtio queue size has been increased from default value of 256 to 1024
  descriptors.
- Adjusted Linux kernel CFS scheduler settings, as detailed on this CSIT wiki
  page: https://wiki.fd.io/view/CSIT/csit-perf-env-tuning-ubuntu1604.

Adjusted Linux kernel CFS settings make the NDR and PDR throughput performance
of VPP+VM system less sensitive to other Linux OS system tasks by reducing
their interference on CPU cores that are designated for critical software
tasks under test, namely VPP worker threads in host and Testpmd threads in
guest dealing with data plan.

Methodology: IPSec with Intel QAT HW cards
------------------------------------------

VPP IPSec performance tests are using DPDK cryptodev device driver in
combination with HW cryptodev devices - Intel QAT 8950 50G - present in
LF FD.io physical testbeds. DPDK cryptodev can be used for all IPSec
data plane functions supported by VPP.

Currently CSIT |release| implements following IPSec test cases:

- AES-GCM, CBC-SHA1 ciphers, in combination with IPv4 routed-forwarding
  with Intel xl710 NIC.
- CBC-SHA1 ciphers, in combination with LISP-GPE overlay tunneling for
  IPv4-over-IPv4 with Intel xl710 NIC.

Methodology: TRex Traffic Generator Usage
-----------------------------------------

The `TRex traffic generator <https://wiki.fd.io/view/TRex>`_ is used for all
CSIT performance tests. TRex stateless mode is used to measure NDR and PDR
throughputs using binary search (NDR and PDR discovery tests) and for quick
checks of DUT performance against the reference NDRs (NDR check tests) for
specific configuration.

TRex is installed and run on the TG compute node. The typical procedure is:

- If the TRex is not already installed on TG, it is installed in the
  suite setup phase - see `TRex intallation`_.
- TRex configuration is set in its configuration file
  ::

  /etc/trex_cfg.yaml

- TRex is started in the background mode
  ::

  $ sh -c 'cd /opt/trex-core-2.25/scripts/ && sudo nohup ./t-rex-64 -i -c 7 --iom 0 > /dev/null 2>&1 &' > /dev/null

- There are traffic streams dynamically prepared for each test. The traffic
  is sent and the statistics obtained using trex_stl_lib.api.STLClient.

**Measuring packet loss**

- Create an instance of STLClient
- Connect to the client
- Add all streams
- Clear statistics
- Send the traffic for defined time
- Get the statistics

If there is a warm-up phase required, the traffic is sent also before test and
the statistics are ignored.

**Measuring latency**

If measurement of latency is requested, two more packet streams are created (one
for each direction) with TRex flow_stats parameter set to STLFlowLatencyStats. In
that case, returned statistics will also include min/avg/max latency values.