Table of Contents
Introduction
As part of the collaboration between VSPERF and ETSI NFV TST working group, we developed and tested new methods to characterize the limits of data plane resources. The new methods are comprised of new search algorithm(s) to discover true resource limits (RFC 2544 Throughput, for example) while avoiding the effects of transient loss in the NFVI platform. The current search algorithm is called Binary Search with Loss Verification. A key aspect of each measurement completed during the search (a trial) is the duration of the test. The Trial Duration must be considerably less than the average time interval between transient events that cause frame loss. The details are described in ETSI GS NFV-TST 009 .
The Long Duration Testing compliments the Binary Search with Loss Verification algorithm by helping to determine (or confirm) the trial duration, and to identify the processes and/or interrupts that cause transient loss. Once the sources of transients have been identified, they can be optimized (for example, by reducing their rate of occurrence) or mitigated (by increasing strategic buffer sizes to absorb more frames during intervals of suspended forwarding).
Testing at Gambia Plugfest
We will be using an expanded set of traffic generators for Long Duration Testing, and primarily iPerf3 during the Plugfest. This means we will use VSPERF in --mode trafficgen-off and operating iPerf3 manually, developing tools to automate data collection as we go.
...
Loss characteristics are a topic for further investigation, and likely expansion of the ETSI NFV specification TST009, clause 11.
First tests with PROX monitoring
As a first step, PROX was started on Node 4 after reboot, with nothing else running:
...
and PROX interrupts on Core 5 (two observed over the 60 second test)
Next steps:
- Obtain second-by-second output from PROX, and confirm Core occupation of OVS-Vanilla, etc.
- OVS-DPDK with isolcpu rcu_nocbs
- and taskset the iPerf process to qa core with no interrupts.
iPerf3 example Output
As an example of the iPerf3 loss counting capability, we have the results of one 60 second test below (where each second would be a trial):
...
This particular test exhibited 6 seconds (Trials) with loss during a 60 second Test, or one loss event every 10 seconds on average. However, many of the Trials with loss occur in a pattern of Loss, no-Loss, Loss over three consecutive Trials. Loss-free intervals were 9, 15, 17, 3, and 8 seconds in length. This testing was conducted at about 70% of the RFC2544 Throughput level.
Related papers:
https://www.net.in.tum.de/fileadmin/bibtex/publications/papers/SPECTS15NAPIoptimization.pdf
...
PROX config:
cd ./samplevnf/VNFs/DPPD-PROX/build
...