Bottleneck in the Network Kit

[waddlesplash]

This is hrev47102. When using Web+ (libnetapi) to download a file over LAN (guest-to-host in VirtualBox, to be precise) the transfer runs at 2.5MB/s. When using "wget" with the exact same config, it runs at 5MB/s.

I did not determine if the bottleneck is always 1/2 speed or if it simply caps at 2.5MB/s. I did notice the CPU is pegged for both libnetapi and wget.

[tqh]

Perhaps you can look at it with Scheduling Recorder? Here is a short howto ​https://plus.google.com/116713839708821747722/posts/QpUzBRczywm

[Coldfirex]

Any change if you give Haiku more than 1 cpu core? Which nic was it given in vbox?

[pulkomandy]

The libnetapi currently uses a thread for each request, which copies network buffers from the network card to BMessages. This design is suboptimal, unfortunately I'm not sure it's possible to do anything better for Web+, as WebKit is designed around processing everything in the BApplication thread, so network data must come as BMessages.

[waddlesplash]

so network data must come as BMessages.

Some sort of a chunk-based ring buffer would make more sense here, I think, with perhaps only periodic wakeups occuring via BMessages. That should avoid a lot of overhead and thus the high CPU usage.

[diver]

Is this fixed with hrev54968?

[pulkomandy]

The libnetapi part should be fixed. Still, 100% CPU use at 5MB/s seems a bit disappointing? Should we continue investigating that?

[diver]

Sure :)