|
From: | Bill Auerbach |
Subject: | [lwip-users] Which opt.h options effect TCP receive throughput? |
Date: | Tue, 22 Jan 2008 14:09:31 -0500 |
Hello, I’m still benchmarking lwIP TCP
(PPC Power Arch) while receiving 2MB files to the PPC. I’m using
raw api (call backs) and NO_SYS. 98% of the processor is processing lwIP
timers and checking for Ethernet packets and the callbacks are doing the work.
I have not moved to Jumbo (Huge) frames yet – I want to get standard
packets as fast as possible. Interestingly, copying from the pbufs to
contiguous memory is using about 10% of the bandwidth *including* using DMA to move the data (it
was 40% with memcpy and 20% with an optimized memcpy). Right now I’m at 464MbS for a 2MB
transfer. Finally the question of the subject line: Which lwIP options
most effect receive throughput? Is it just these, which I’ve set as
follows: #define
TCP_WND
(64*1024-1) #define
TCP_MSS
(8*1024) One thing I notice is there is an ACK sent
for every 2 packets received. Does this sound right? (2874
received, 1437 sent). I tried a direct point to point link (PC
to PPC box) with no change in the results. So the router is giving this
burst everything – in fact I see no intervening packets in wireshark from
the first to last TCP packet in my transfer (this surprised me actually). What is the theoretical limit with
TCP? Is it the PC slowing things down, or am I simply close to “as
good as it gets”? Thanks, Bill Auerbach |
[Prev in Thread] | Current Thread | [Next in Thread] |