| On February 3rd, we conducted a very successful 2Gbit/s e-VLBI test. Many EVN stations participated; Effelsberg, Hartebeesthoek, Medicina, Noto, Onsala and Yebes all streamed data from their FILA10G boards at 2 Gbit/s while Torun used the traditional e-VLBI method with a Mark5B at 1 Gbit/s. We started our test by streaming "default" VDIF frames with a payload of 8000 bytes with multiple channels packed into a single frame (single-threaded, multi-channel). This worked really well (upon the first attempt!) thanks to recent optimizations in the SFXC software correlator and tuning of the Infiniband network of the correlator compute cluster. The total data rate was in excess of 13.5 Gbit/s, and it ran for more than an hour without any hiccups and with minimal packets loss. There were fringes to all participating stations. We pretty much maxed out one of our 10 Gbit/s links to SURFNet while doing this!
We also tested VDIF frames with a payload of 2000 bytes using separate frames for each channel (multi-threaded, single-channel), where the FILA10G does the corner-turning. This took a bit more effort, but eventually we succeeded. It's interesting to note that while letting the FILA10G do the corner-turning saves some computation in the correlator, this is pretty much offset by the increased per-packet overhead as we need to process 4 times as many packets per second. The overhead of using smaller frames is nicely visible in the network throughput graph.
This test proves that the SFXC correlator at JIVE can comfortably handle real-time correlation of 2 Gbit/s for six-and-a-half stations. There is probably enough spare capacity for adding one or two more stations as the cluster nodes were still partly idle and not all nodes were used for this test. There are still a number of technical issues to sort out to make sure the data streams can be started and stopped on demand. But we're getting really close to offering 2Gbit/s e-VLBI to our users.
We'd like to thank the VLBI friends at the participating stations for their help with this test.
Editor's note: It is not necessary to understand all the jargon to be really impressed with this stuff.