Improving WinDivert perfomance by using BatchMode offered in WinDivertRecvEx

[trudyhood]

I have found out that WinDivert performance can be increased significantly by using batch mode offered in WinDivertRecvEx. Can you implement WinDivertDevice with BatchMode using WinDivertRecvEx instead WinDivertRecv? The performance is near 3x on high bandwidth and CPU utilization looks lower.

Please consider WinDivertSendEx too.

[trudyhood]

Sorry, no any comment about this?

[chmorgan]

It’s a good idea. Let me look at how we might support that feature once I’ve wrapped up some changes I have pending to clean up the API.

…

On Sun, May 2, 2021 at 7:31 PM trudyhood ***@***.***> wrote: Sorry, no any comment about this? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#273 (comment)>, or unsubscribe <http://github.com/notifications/unsubscribe-auth/AAJH4AEXIIPEBKHNVPN5TZDTLXODTANCNFSM43ASH5UA> .

[trudyhood]

Very good, really waiting for that.
Thank you

[chmorgan]

Alright, took a look at the recv api. I'm not sure if/when I'd get to the changes but if this was implemented I think we'd want to do something like:

• Add a new WinDivert specific GetNext() that takes in an array of 'out PacketCapture[] e' (if this is possible)
• We need to have WinDivertDevice support an array of 'buffer's, likely with some dynamic allocation such that we allocate and grow them based on calls to GetNext() asking for multiple packets so we don't take up as much memory until the first call to get X packets, for example.
• GetNext() would check to see how many entries were in the PacketCapture[] the array and call WinDivert to receive the min of e.length(), and the array of internal buffers.
• And GetNext() would return the number of packets captured.

and of course we'd want to add a test to test the new method.

[trudyhood]

I hope you make the changes. I should say congratulation for updating to version 6. It looks version 6 is 30% faster. Using batch mode with WinDivert increases the speed by more than 300%!
You can check the performance difference with windivert passthru.exe on a gigabit network.

passthru.exe [filter] [num-threads] [batch-size]
passthru.exe true 1 10

[kayoub5]

Keeping this open until WinDivertSendEx is supported as well

[verdie-g]

DeviceModes.MaxResponsiveness is currently unused for WinDivertDevice.

Would it make sense to read X packets in the first call to GetNextPacket then returns the already loaded packets on the next calls. But if DeviceModes.MaxResponsiveness is set, then the batch size becomes 1.

[kayoub5]

That generally have little benefit, the main point of batch processing for a network processing library is being able to process the batch elements in parallel at the application side.