diff options
Diffstat (limited to 'docs/results/os-onos-sfc-ha.rst')
-rw-r--r-- | docs/results/os-onos-sfc-ha.rst | 517 |
1 files changed, 517 insertions, 0 deletions
diff --git a/docs/results/os-onos-sfc-ha.rst b/docs/results/os-onos-sfc-ha.rst new file mode 100644 index 000000000..e52ae3d55 --- /dev/null +++ b/docs/results/os-onos-sfc-ha.rst @@ -0,0 +1,517 @@ +.. This work is licensed under a Creative Commons Attribution 4.0 International +.. License. +.. http://creativecommons.org/licenses/by/4.0 + + +=============================== +Test Results for os-onos-sfc-ha +=============================== + +.. toctree:: + :maxdepth: 2 + + +fuel +==== + +.. _Grafana: http://testresults.opnfv.org/grafana/dashboard/db/yardstick-main +.. _POD2: https://wiki.opnfv.org/pharos?&#community_test_labs + +Overview of test results +------------------------ + +See Grafana_ for viewing test result metrics for each respective test case. It +is possible to chose which specific scenarios to look at, and then to zoom in +on the details of each run test scenario as well. + +All of the test case results below are based on 4 scenario test runs, each run +on the Ericsson POD2_ or LF POD2_ between September 5 and 10 in 2016. + +TC002 +----- +The round-trip-time (RTT) between 2 VMs on different blades is measured using +ping. Most test run measurements result on average between 0.5 and 0.6 ms. +A few runs start with a 1 - 1.5 ms RTT spike (This could be because of normal ARP +handling). One test run has a greater RTT spike of 1.9 ms, which is the same +one with the 0.7 ms average. The other runs have no similar spike at all. +To be able to draw conclusions more runs should be made. +SLA set to 10 ms. The SLA value is used as a reference, it has not +been defined by OPNFV. + +TC005 +----- +The IO read bandwidth looks similar between different dates, with an +average between approx. 170 and 200 MB/s. Within each test run the results +vary, with a minimum 2 MB/s and maximum 838 MB/s on the totality. Most runs +have a minimum BW of 3 MB/s (two runs at 2 MB/s). The maximum BW varies more in +absolute numbers between the dates, between 617 and 838 MB/s. +SLA set to 400 MB/s. The SLA value is used as a reference, it has not been +defined by OPNFV. + +TC010 +----- +The measurements for memory latency are similar between test dates and result +in approx. 1.2 ns. The variations within each test run are similar, between +1.215 and 1.219 ns. One exception is February 16, where the average is 1.222 +and varies between 1.22 and 1.28 ns. +SLA set to 30 ns. The SLA value is used as a reference, it has not been defined +by OPNFV. + +TC011 +----- +Packet delay variation between 2 VMs on different blades is measured using +Iperf3. On the first date the reported packet delay variation varies between +0.0025 and 0.011 ms, with an average delay variation of 0.0067 ms. +On the second date the delay variation varies between 0.002 and 0.006 ms, with +an average delay variation of 0.004 ms. + +TC012 +----- +Between test dates, the average measurements for memory bandwidth vary between +17.4 and 17.9 GB/s. Within each test run the results vary more, with a minimal +BW of 16.4 GB/s and maximum of 18.2 GB/s on the totality. +SLA set to 15 GB/s. The SLA value is used as a reference, it has not been +defined by OPNFV. + +TC014 +----- +The Unixbench processor test run results vary between scores 3080 and 3240, +one result each date. The average score on the total is 3150. +No SLA set. + +TC037 +----- +The amount of packets per second (PPS) and round trip times (RTT) between 2 VMs +on different blades are measured when increasing the amount of UDP flows sent +between the VMs using pktgen as packet generator tool. + +Round trip times and packet throughput between VMs can typically be affected by +the amount of flows set up and result in higher RTT and less PPS throughput. + +The RTT results are similar throughout the different test dates and runs at +approx. 15 ms. Some test runs show an increase with many flows, in the range +towards 16 to 17 ms. One exception standing out is Feb. 15 where the average +RTT is stable at approx. 13 ms. The PPS results are not as consistent as the +RTT results. +In some test runs when running with less than approx. 10000 flows the PPS +throughput is normally flatter compared to when running with more flows, after +which the PPS throughput decreases. Around 20 percent decrease in the worst +case. For the other test runs there is however no significant change to the PPS +throughput when the number of flows are increased. In some test runs the PPS +is also greater with 1000000 flows compared to other test runs where the PPS +result is less with only 2 flows. + +The average PPS throughput in the different runs varies between 414000 and +452000 PPS. The total amount of packets in each test run is approx. 7500000 to +8200000 packets. One test run Feb. 15 sticks out with a PPS average of +558000 and approx. 1100000 packets in total (same as the on mentioned earlier +for RTT results). + +There are lost packets reported in most of the test runs. There is no observed +correlation between the amount of flows and the amount of lost packets. +The lost amount of packets normally range between 100 and 1000 per test run, +but there are spikes in the range of 10000 lost packets as well, and even +more in a rare cases. + +CPU utilization statistics are collected during UDP flows sent between the VMs +using pktgen as packet generator tool. The average measurements for CPU +utilization ratio vary between 1% to 2%. The peak of CPU utilization ratio +appears around 7%. + +TC069 +----- +Between test dates, the average measurements for memory bandwidth vary between +15.5 and 25.4 GB/s. Within each test run the results vary more, with a minimal +BW of 9.7 GB/s and maximum of 29.5 GB/s on the totality. +SLA set to 6 GB/s. The SLA value is used as a reference, it has not been +defined by OPNFV. + +TC070 +----- +The amount of packets per second (PPS) and round trip times (RTT) between 2 VMs +on different blades are measured when increasing the amount of UDP flows sent +between the VMs using pktgen as packet generator tool. + +Round trip times and packet throughput between VMs can typically be affected by +the amount of flows set up and result in higher RTT and less PPS throughput. + +The RTT results are similar throughout the different test dates and runs at +approx. 15 ms. Some test runs show an increase with many flows, in the range +towards 16 to 17 ms. One exception standing out is Feb. 15 where the average +RTT is stable at approx. 13 ms. The PPS results are not as consistent as the +RTT results. +In some test runs when running with less than approx. 10000 flows the PPS +throughput is normally flatter compared to when running with more flows, after +which the PPS throughput decreases. Around 20 percent decrease in the worst +case. For the other test runs there is however no significant change to the PPS +throughput when the number of flows are increased. In some test runs the PPS +is also greater with 1000000 flows compared to other test runs where the PPS +result is less with only 2 flows. + +The average PPS throughput in the different runs varies between 414000 and +452000 PPS. The total amount of packets in each test run is approx. 7500000 to +8200000 packets. One test run Feb. 15 sticks out with a PPS average of +558000 and approx. 1100000 packets in total (same as the on mentioned earlier +for RTT results). + +There are lost packets reported in most of the test runs. There is no observed +correlation between the amount of flows and the amount of lost packets. +The lost amount of packets normally range between 100 and 1000 per test run, +but there are spikes in the range of 10000 lost packets as well, and even +more in a rare cases. + +Memory utilization statistics are collected during UDP flows sent between the +VMs using pktgen as packet generator tool. The average measurements for memory +utilization vary between 225MB to 246MB. The peak of memory utilization appears +around 340MB. + +TC071 +----- +The amount of packets per second (PPS) and round trip times (RTT) between 2 VMs +on different blades are measured when increasing the amount of UDP flows sent +between the VMs using pktgen as packet generator tool. + +Round trip times and packet throughput between VMs can typically be affected by +the amount of flows set up and result in higher RTT and less PPS throughput. + +The RTT results are similar throughout the different test dates and runs at +approx. 15 ms. Some test runs show an increase with many flows, in the range +towards 16 to 17 ms. One exception standing out is Feb. 15 where the average +RTT is stable at approx. 13 ms. The PPS results are not as consistent as the +RTT results. +In some test runs when running with less than approx. 10000 flows the PPS +throughput is normally flatter compared to when running with more flows, after +which the PPS throughput decreases. Around 20 percent decrease in the worst +case. For the other test runs there is however no significant change to the PPS +throughput when the number of flows are increased. In some test runs the PPS +is also greater with 1000000 flows compared to other test runs where the PPS +result is less with only 2 flows. + +The average PPS throughput in the different runs varies between 414000 and +452000 PPS. The total amount of packets in each test run is approx. 7500000 to +8200000 packets. One test run Feb. 15 sticks out with a PPS average of +558000 and approx. 1100000 packets in total (same as the on mentioned earlier +for RTT results). + +There are lost packets reported in most of the test runs. There is no observed +correlation between the amount of flows and the amount of lost packets. +The lost amount of packets normally range between 100 and 1000 per test run, +but there are spikes in the range of 10000 lost packets as well, and even +more in a rare cases. + +Cache utilization statistics are collected during UDP flows sent between the +VMs using pktgen as packet generator tool. The average measurements for cache +utilization vary between 205MB to 212MB. + +TC072 +----- +The amount of packets per second (PPS) and round trip times (RTT) between 2 VMs +on different blades are measured when increasing the amount of UDP flows sent +between the VMs using pktgen as packet generator tool. + +Round trip times and packet throughput between VMs can typically be affected by +the amount of flows set up and result in higher RTT and less PPS throughput. + +The RTT results are similar throughout the different test dates and runs at +approx. 15 ms. Some test runs show an increase with many flows, in the range +towards 16 to 17 ms. One exception standing out is Feb. 15 where the average +RTT is stable at approx. 13 ms. The PPS results are not as consistent as the +RTT results. +In some test runs when running with less than approx. 10000 flows the PPS +throughput is normally flatter compared to when running with more flows, after +which the PPS throughput decreases. Around 20 percent decrease in the worst +case. For the other test runs there is however no significant change to the PPS +throughput when the number of flows are increased. In some test runs the PPS +is also greater with 1000000 flows compared to other test runs where the PPS +result is less with only 2 flows. + +The average PPS throughput in the different runs varies between 414000 and +452000 PPS. The total amount of packets in each test run is approx. 7500000 to +8200000 packets. One test run Feb. 15 sticks out with a PPS average of +558000 and approx. 1100000 packets in total (same as the on mentioned earlier +for RTT results). + +There are lost packets reported in most of the test runs. There is no observed +correlation between the amount of flows and the amount of lost packets. +The lost amount of packets normally range between 100 and 1000 per test run, +but there are spikes in the range of 10000 lost packets as well, and even +more in a rare cases. + +Network utilization statistics are collected during UDP flows sent between the +VMs using pktgen as packet generator tool. Total number of packets received per +second was average on 200 kpps and total number of packets transmitted per +second was average on 600 kpps. + +Detailed test results +--------------------- +The scenario was run on Ericsson POD2_ and LF POD2_ with: +Fuel 9.0 +OpenStack Mitaka +Onos Goldeneye +OpenVirtualSwitch 2.5.90 +OpenDayLight Beryllium + +Rationale for decisions +----------------------- +Pass + +Tests were successfully executed and metrics collected. +No SLA was verified. To be decided on in next release of OPNFV. + +Conclusions and recommendations +------------------------------- +The pktgen test configuration has a relatively large base effect on RTT in +TC037 compared to TC002, where there is no background load at all. Approx. +15 ms compared to approx. 0.5 ms, which is more than a 3000 percentage +difference in RTT results. +Especially RTT and throughput come out with better results than for instance +the *fuel-os-nosdn-nofeature-ha* scenario does. The reason for this should +probably be further analyzed and understood. Also of interest could be +to make further analyzes to find patterns and reasons for lost traffic. +Also of interest could be to see if there are continuous variations where +some test cases stand out with better or worse results than the general test +case. + + +Joid +===== + +.. _Grafana: http://testresults.opnfv.org/grafana/dashboard/db/yardstick-main +.. _POD6: https://wiki.opnfv.org/pharos?&#community_test_labs + +Overview of test results +------------------------ + +See Grafana_ for viewing test result metrics for each respective test case. It +is possible to chose which specific scenarios to look at, and then to zoom in +on the details of each run test scenario as well. + +All of the test case results below are based on 4 scenario test runs, each run +on the Intel POD6_ between September 8 and 11 in 2016. + +TC002 +----- +The round-trip-time (RTT) between 2 VMs on different blades is measured using +ping. Most test run measurements result on average between 1.35 ms and 1.57 ms. +Only one test run has reached greatest RTT spike of 2.58 ms. Meanwhile, the +smallest network latency is 1.11 ms, which is obtained on Sep. 11st. In +general, the average of network latency of the four test runs are between 1.35 +ms and 1.57 ms. SLA set to be 10 ms. The SLA value is used as a reference, it +has not been defined by OPNFV. + +TC005 +----- +The IO read bandwidth actually refers to the storage throughput, which is +measured by fio and the greatest IO read bandwidth of the four runs is 175.4 +MB/s. The IO read bandwidth of the three runs looks similar, with an average +between 43.7 and 56.3 MB/s, except one on Sep. 8, for its maximum storage +throughput is only 107.9 MB/s. One of the runs has a minimum BW of 478 KM/s and +other has a maximum BW of 168.6 MB/s. The SLA of read bandwidth sets to be +400 MB/s, which is used as a reference, and it has not been defined by OPNFV. + +The results of storage IOPS for the four runs look similar with each other. The +IO read times per second of the four test runs have an average value between +978 per second and 1.20 K/s, and meanwhile, the minimum result is only 36 times +per second. + +TC010 +----- +The tool we use to measure memory read latency is lmbench, which is a series of +micro benchmarks intended to measure basic operating system and hardware system +metrics. The memory read latency of the four runs is between 1.164 ns and 1.244 +ns on average. The variations within each test run are quite different, some +vary from a large range and others have a small change. For example, the +largest change is on September 10, the memory read latency of which is ranging +from 1.128 ns to 1.381 ns. However, the results on September 11 change very +little. The SLA sets to be 30 ns. The SLA value is used as a reference, it has +not been defined by OPNFV. + +TC011 +----- +Iperf3 is a tool for evaluating the packet delay variation between 2 VMs on +different blades. The reported packet delay variations of the four test runs +differ from each other. In general, the packet delay of two runs look similar, +for they both stay stable within each run. And the mean packet delay of them +are 0.0772 ms and 0.0788 ms respectively. Of the four runs, the fourth has the +worst result, because the packet delay reaches 0.0838 ms. The rest one has a +large wide range from 0.0666 ms to 0.0798 ms. The SLA value sets to be 10 ms. +The SLA value is used as a reference, it has not been defined by OPNFV. + +TC012 +----- +Lmbench is also used to measure the memory read and write bandwidth, in which +we use bw_mem to obtain the results. Among the four test runs, the trend of the +memory bandwidth almost look similar, which all have a large wide range, and +the minimum and maximum results are 9.02 GB/s and 18.14 GB/s. Here SLA set to +be 15 GB/s. The SLA value is used as a reference, it has not been defined by +OPNFV. + +TC014 +----- +The Unixbench is used to evaluate the IaaS processing speed with regards to +score of single cpu running and parallel running. It can be seen from the +dashboard that the processing test results vary from scores 3395 to 3475, and +there is only one result one date. No SLA set. + +TC037 +----- +The amount of packets per second (PPS) and round trip times (RTT) between 2 VMs +on different blades are measured when increasing the amount of UDP flows sent +between the VMs using pktgen as packet generator tool. + +Round trip times and packet throughput between VMs can typically be affected by +the amount of flows set up and result in higher RTT and less PPS throughput. + +The mean packet throughput of the four test runs is between 362.1 kpps and +363.5 kpps, of which the result of the third run is the highest. The RTT +results of all the test runs keep flat at approx. 17 ms. It is obvious that the +PPS results are not as consistent as the RTT results. + +The No. flows of the four test runs are 240 k on average and the PPS results +look a little waved since the largest packet throughput is 418.1 kpps and the +minimum throughput is 326.5 kpps respectively. + +There are no errors of packets received in the four runs, but there are still +lost packets in all the test runs. The RTT values obtained by ping of the four +runs have the similar average vaue, that is approx. 17 ms, of which the worst +RTT is 39 ms on Sep. 11st. + +CPU load is measured by mpstat, and CPU load of the four test runs seem a +little similar, since the minimum value and the peak of CPU load is between 0 +percent and nine percent respectively. And the best result is obtained on Sep. +10, with an CPU load of nine percent. + +TC069 +----- +With the block size changing from 1 kb to 512 kb, the memory write bandwidth +tends to become larger first and then smaller within every run test, which +rangs from 25.9 GB/s to 26.6 GB/s and then to 18.1 GB/s on average. Since the +test id is one, it is that only the INT memory write bandwidth is tested. On +the whole, when the block size is from 2 kb to 16 kb, the memory write +bandwidth look similar with a minimal BW of 22.1 GB/s and peak value of 28.6 +GB/s. And then with the block size becoming larger, the memory write bandwidth +tends to decrease. SLA sets to be 7 GB/s. The SLA value is used as a reference, +it has not been defined by OPNFV. + +TC070 +----- +The amount of packets per second (PPS) and round trip times (RTT) between 2 VMs +on different blades are measured when increasing the amount of UDP flows sent +between the VMs using pktgen as packet generator tool. + +Round trip times and packet throughput between VMs can typically be affected by +the amount of flows set up and result in higher RTT and less PPS throughput. + +The network latency is measured by ping, and the results of the four test runs +look similar with each other, and within these test runs, the maximum RTT can +reach 39 ms and the average RTT is usually approx. 17 ms. The network latency +tested on Sep. 11 shows that it has a peak latency of 39 ms. But on the whole, +the average RTTs of the five runs keep flat and the network latency is +relatively short. + +Memory utilization is measured by free, which can display amount of free and +used memory in the system. The largest amount of used memory is 270 MiB on the +first two runs. In general, the mean used memory of two test runs have very +large memory utilization, which can reach 264 MiB on average. And the other two +runs have a large wide range of memory usage with the minimum value of 150 MiB +and the maximum value of 270 MiB. On the other hand, for the mean free memory, +the four test runs have the similar trend with that of the mean used memory. +In general, the mean free memory change from 220 MiB to 342 MiB. + +Packet throughput and packet loss can be measured by pktgen, which is a tool +in the network for generating traffic loads for network experiments. The mean +packet throughput of the four test runs seem quite different, ranging from +326.5 kpps to 418.1 kpps. The average number of flows in these tests is +240000, and each run has a minimum number of flows of 2 and a maximum number +of flows of 1.001 Mil. At the same time, the corresponding packet throughput +differ between 326.5 kpps and 418.1 kpps with an average packet throughput between +361.7 kpps and 363.5 kpps. In summary, the PPS results seem consistent. Within each +test run of the four runs, when number of flows becomes larger, the packet +throughput seems not larger at the same time. + +TC071 +----- +The amount of packets per second (PPS) and round trip times (RTT) between 2 VMs +on different blades are measured when increasing the amount of UDP flows sent +between the VMs using pktgen as packet generator tool. + +Round trip times and packet throughput between VMs can typically be affected by +the amount of flows set up and result in higher RTT and less PPS throughput. + +The network latency is measured by ping, and the results of the four test runs +look similar with each other. Within each test run, the maximum RTT is only 47 +ms and the average RTT is usually approx. 15 ms. On the whole, the average +RTTs of the four runs keep stable and the network latency is relatively small. + +Cache utilization is measured by cachestat, which can display size of cache and +buffer in the system. Cache utilization statistics are collected during UDP +flows sent between the VMs using pktgen as packet generator tool. The largest +cache size is 214 MiB, which is same for the four runs, and the smallest cache +size is 94 MiB. On the whole, the average cache size of the four runs look the +same and is between 198 MiB and 207 MiB. Meanwhile, the tread of the buffer +size keep flat, since they have a minimum value of 7 MiB and a maximum value of +8 MiB, with an average value of about 7.9 MiB. + +Packet throughput can be measured by pktgen, which is a tool in the network for +generating traffic loads for network experiments. The mean packet throughput of +the four test runs seem quite the same, which is approx. 363 kpps. The average +number of flows in these tests is 240k, and each run has a minimum number of +flows of 2 and a maximum number of flows of 1.001 Mil. At the same time, the +corresponding packet throughput differ between 327 kpps and 418 kpps with an +average packet throughput of about 363 kpps. Within each test run of the four +runs, when number of flows becomes larger, the packet throughput seems not +larger in the meantime. + +TC072 +----- +The amount of packets per second (PPS) and round trip times (RTT) between 2 VMs +on different blades are measured when increasing the amount of UDP flows sent +between the VMs using pktgen as packet generator tool. + +Round trip times and packet throughput between VMs can typically be affected by +the amount of flows set up and result in higher RTT and less PPS throughput. + +The RTT results are similar throughout the different test dates and runs +between 0 ms and 47 ms with an average leatency of less than 16 ms. The PPS +results are not as consistent as the RTT results, for the mean packet +throughput of the four runs differ from 361.7 kpps to 365.0 kpps. + +Network utilization is measured by sar, that is system activity reporter, which +can display the average statistics for the time since the system was started. +Network utilization statistics are collected during UDP flows sent between the +VMs using pktgen as packet generator tool. The largest total number of packets +transmitted per second look similar for two test runs, whose values change a +lot from 10 pps to 432 kpps. While results of the other test runs seem the same +and keep stable with the average number of packets transmitted per second of 10 +pps. However, the total number of packets received per second of the four runs +look similar, which have a large wide range of 2 pps to 657 kpps. + +In some test runs when running with less than approx. 250000 flows the PPS +throughput is normally flatter compared to when running with more flows, after +which the PPS throughput decreases. For the other test runs there is however no +significant change to the PPS throughput when the number of flows are +increased. In some test runs the PPS is also greater with 250000 flows +compared to other test runs where the PPS result is less with only 2 flows. + +There are lost packets reported in most of the test runs. There is no observed +correlation between the amount of flows and the amount of lost packets. +The lost amount of packets normally differs a lot per test run. + +Detailed test results +--------------------- +The scenario was run on Intel POD6_ with: +Joid +OpenStack Mitaka +Onos Goldeneye +OpenVirtualSwitch 2.5.90 +OpenDayLight Beryllium + +Rationale for decisions +----------------------- +Pass + +Conclusions and recommendations +------------------------------- +Tests were successfully executed and metrics collected. +No SLA was verified. To be decided on in next release of OPNFV. + |