←back to thread

193 points lijunhao | 1 comments | | HN request time: 0.21s | source
Show context
jnck ◴[] No.42155820[source]
Great. Now we could gain detailed insights into how our system is behaving in real time, which is invaluable for troubleshooting and optimizing performance. For those who just heard eBPF, there is the fun-damental source about it [0].

Links: [0]: https://ebpf.io/books/buzzing-across-space-illustrated-child...

replies(3): >>42156293 #>>42156344 #>>42156769 #
burnt-resistor ◴[] No.42156344[source]
Nice nice!

Btw, I'm wondering if OFED and/or DPDK are also still used, and if they're still used for fast packet pushing.

replies(4): >>42156381 #>>42156410 #>>42156915 #>>42159278 #
1. jpgvm ◴[] No.42156915[source]
By OFED I assume you are meaning RDMA and yes, it's used extensively. Not just in HPC but anywhere you are doing high performance collective communication. Frameworks like MPI, UPC/UPC++, NCCL, UCX etc are all underpinned by RDMA. Most of the AI distributed training frameworks are MPI based for example.

OFED is less of a thing now because most of the work has gone upstream, both into the kernel and into the rdma-core userland.

Also worth mentioning that MLNX_OFED (sometimes called MOFED) is now being transitioned into DOCA-Host. This is mostly because of that aforementioned upstreaming and the move towards more SmartNIC stuff (ala Bluefield) being the focus as core RDMA support is mostly provided by upstream.