Repository navigation
Question: UDP multicast receive scaling, RSS queue steering, and AF_XDP applicability on Windows #1053
Unanswered
Roberto Musso (Rob-NDI)
asked this question in
Q&A
Replies: 1 comment
|
Hey!
In general, Windows has more optimizations for unicast traffic than multicast. Out of curiosity, what packet rate and total bitrate is causing this to stall? Is the lack of spreading occurring all within an IP 2-tuple or UDP 4-tuple? Modern server-grade NICs are expected to hash and spread UDP 4-tuples across CPUs, so distinct sender/receiver pairs should get CPU spreading. I'm not sure if you've captured and taken a look at CPU flame graphs to find any unexpected hot spots in the kernel receive path. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hello,
I am investigating a high-bitrate UDP multicast receive workload on Windows, related to professional media/video-over-IP workflows.
At the network level, multicast behaves correctly: IGMP Snooping and Querier are configured, the sender transmits a single copy of each stream, and receiver NIC bandwidth does not increase when multiple local receivers subscribe to the same multicast source.
The issue appears on the Windows receiver side. With multiple UDP multicast streams, or multiple local receivers consuming the same multicast source, the receive-side workload can concentrate on a single logical CPU core. When that core reaches 100%, packet/video drops occur, even though total CPU and NIC bandwidth are still available.
Windows counters show that the saturated core is dominated by Privileged Time, not User Time. DPC and Interrupt Time are relatively low. Process Explorer shows hot activity in the System process, associated with ndis.sys.
We reproduced similar behavior across multiple NIC families, including Intel X520/X710, Intel E810-XXV-2, Mellanox ConnectX-4 Lx, and Broadcom BCM57810. Changing NICs and RSS settings changes the exact core mapping behavior, but does not eliminate the issue. Linux on the same hardware and same network configuration distributes the workload much better.
My questions are:
Is there any supported mechanism on Windows to explicitly steer UDP multicast flows to different RSS queues or CPU cores?
Can XDP for Windows or AF_XDP be used to handle this kind of high-bitrate UDP multicast receive workload more efficiently than the standard NDIS/UDP socket path?
Is there any Windows equivalent to Linux ethtool ntuple / Intel Flow Director style filtering, where specific UDP multicast flows can be mapped to specific receive queues?
If multiple local sockets/processes subscribe to the same multicast flow, is the receive-side delivery expected to remain tied to the same RSS queue/core?
Any guidance on whether XDP for Windows is relevant to this use case would be greatly appreciated.
All reactions