|
From: | Zayaz Volk |
Subject: | Re: [lwip-users] LWIP/WIN32 UDP performance (throughtput) |
Date: | Tue, 7 Feb 2012 16:18:17 +0000 |
Hi,
Thanks for your answers. Achieving 500-600Mbs using TCP on dedicated hardware is quite promising, so it might be some inefficiencies related to windows port or wincap (although Ethereal, i.e. Wireshark is using the wincap library for its packet sniffing, so i would guess it shouldn't degrade the performance by this much (magnitude of 10 vs regular WinSock API, 4-5Mbytes/susing lwip/Win32 vs 40-50Mbytes/s using WinSock)). Meanwhile, I am trying to profile the stack itself, till the pcapif_low_level_output function is being called (it seems to be the transfer point into wincap library by calling the pcap_sendpacket). Again, if somebody has been using the Win32 or Linux ports and has measured the throughput please let me know. I will also look into lwiopt.h to to try and optimize some values there, mostly regarding the memory and buffers constraints. I will post my results and timings later. Any suggestions for "optimized" lwiopt.h values for the systems with more than just few Kbytes of RAM ? Thank you, Stas From: address@hidden To: address@hidden Date: Tue, 7 Feb 2012 08:46:24 -0500 Subject: Re: [lwip-users] LWIP/WIN32 UDP performance (throughtput) I totally agree with Simon, at least with RAW API I cant speak for NETCONN performance. My embedded system is like yours (667Mhz CPU and 256MB 667Mhz DDR2) and is zero-copy with optimized checksum (aligned accesses and in assembly). I can reach 500-600+MbS with TCP. In my testing of UDP the PC in some instances was not able to keep pace receiving. I do not run lwIP on the PC we use Win32 sockets so far without issues.
It would be nice to see a standard Win32/Linux program to communicate with a simple lwIP test program (RAW, NETCONN, sockets) to create a list of lwIP platforms and the results of the tests.
From:
address@hidden
[mailto:address@hidden On Behalf
Of Simon Goldschmidt
Zayaz Volk <address@hidden> wrote:
In my opinion, the win32 port is mainly limited by the netif driver: winpcap doesn't seem to give the best performance, since we have to copy packets twice on RX (and I don't remember if it's once or twice on TX). Also, lwIP doesn't really benefit from multiple CPU cores. Instead, you're right that it is targeted on low resources rather that throughput.
Still, i think lwIP should be able to achieve the performance you want if: A) your netif driver correctly supports zero copy and B) you handle the stack "correctly" so that zero copy is possible (I.e. data passed to a TX function must be left unchanged in memory until the netif driver has actually sent it out).
Simon _______________________________________________ lwip-users mailing list address@hidden https://lists.nongnu.org/mailman/listinfo/lwip-users |
[Prev in Thread] | Current Thread | [Next in Thread] |