For those responsible for knowing and ensuring the status of optical connectivity services in Datacenters: we share recommendations to guarantee the quality of these services.
• Performance Problem in Datacenter Data Links After certifying the internal plant at “Tier 1” level and then “Tier 2” (for these measurements we recommend Viavi OTLS55 and Viavi Certifier10G/40G testers), it often occurs that the service/customer traffic rate between different plant elements does not reach 100%.
• Why Is This Happening? In the case of links between Datacenter equipment, at typical rates of 10G/25G/40G/100G, it often occurs that TCP traffic fails to reach its maximum possible throughput. In these cases, you can have L3 at full rate but a percentage of TCP traffic either in retransmission or idle. On a high-performance and critical traffic circuit, this is definitely harmful.
• What Causes This Problem? The reason for this problem is that TCP throughput depends critically on how TCP flows are “windowed” and “parallelized” (server configurations between endpoints), and what buffer resources and QoS provisions have been deployed in the network (configurations typical of Switches and Routers).
How to Measure It Correctly? If we wish to intervene and run traffic injection tests, we find that it is not easy to emulate realistically with test traffic. A first approach that is typically done is to inject traffic against an IPERF server, which gives us an “estimate” of L4 capacity. But this is not the definitive way to measure, since first you could be adding the problem to the measurement method itself, and second, you cannot test all possible traffic configurations between endpoints until you achieve maximum throughput.
The Test & Measurement industry solved this issue through RFC6349 procedure (which Viavi co-authored with companies such as AT&T). In this recommendation, two instruments are used, one at each end (one as server, the other as client), traffic is injected emulating multiple TCP flows, multiple window sizes, considering latency changes, buffering peaks, retransmissions, etc.
Our instruments implement an automated test flow, first measuring “physical” traffic up to L3 according to RFC2544 standard, then running the ViaviTruespeed test (RFC6349), and finally generating reports of results. With this information, users will be able to discern whether the problem is TCP “windowing”, incorrect “Buffer” handling, excessive latencies, or is associated with L3 traffic conditions, etc. With this information, configuration adjustments are made to the equipment, achieving 100% TCP throughput.
For more information, we recommend the following application notes:
https://www.viavisolutions.com/en-us/literature/data-center-use-case-test-guide-case-studies-en.pdf
https://www.viavisolutions.com/en-us/literature/rfc-6349-testing-truespeed-application-notes-en.pdf
https://www.tecnous.com/wp-content/uploads/2020/05/t-berd-mts-5800-100g-brochure-en-1.pdf
Ing. Emiliano Alaimo
Tecnous SA