perfSONAR

perfSONAR (Performance focused Service Oriented Network monitoring ARchitecture) is an open-source, active network measurement toolkit that provides federated coverage of paths and helps establish end-to-end user expectations.

To identify network problems, it is important to compare active measurements against predefined notions of successful networks. Performance probes are placed in paths of interest, such as campus network endpoints, demarcations between networks, within carrier points of presence, at exchange points, and near data resources such as storage and computing elements.

To provide measurement baselines, some 2000 perfSONAR instances are deployed worldwide, representing around 300 domains, and many of which are available for the open testing of key measures of network performance. Ths global infrastructure helps to identify and isolate problems as they occur, making the role of supporting network users easier for engineering teams, and increasing productivity in the use of network resources.

perfSONAR provides a uniform interface that allows for the scheduling of measurements, storage of data in uniform formats, and scalable methods to retrieve data and generate visualizations. This extensible system can be modified to support new metrics, with a variety of ways to present data.

perfSONAR features

perfSONAR supports the following features and tools:

Active measurement with scheduled tests
Tools: Bwctl (Iperf, iperf3, Nuttcp tests), ping, OWAMP tests (one-way latency). traceroute, tracepath
Visualizations of stored data (MadDash)
Directory service (to find perfSONAR instance around the world)

TPVM deployment considerations

Although SLX-OS allows perfSONAR to run on TPVM, it is recommended that this application be run on a dedicated server to mitigate risks posed by the VM environment, for the following reasons:

Time keeping: Some virtualization environments implement clock management as a function of the hypervisor and VM communication channel, rather than using a stabilization daemon such as NTP. This could result in timing skipping forward or backward, making it generally unpredictable for measurement.
Data path: Additional hypervisor layers can cause undesired latency.
Resource management: Because VMs share physical hardware and might get swapped, this might introduce additional errors in network performance measurements.

Reason (2) is mitigated in TPVM deployments by directly assigning the insight interface to the VM.

Reason (3) can be potentially mitigated by pinning one or more CPU cores to the VM.

Reason (1) can also be mitigated, such as by running NTP between guest and host, but this still not provide sufficient accuracy.

The perSONAR development team has identified several use cases that can work in VM environments, provided the known issues are mitigated. However, the high-speed throughput and OWAMP tests do not perform well.