Re: UDP sending perfomance in Gbit Ethernet
- From: "m" <m@xxx>
- Date: Wed, 21 Dec 2005 12:20:09 -0500
If the throughput changes with network card settings (interupt rate), then
the problem is more lily with the NIC or driver than with the stack.
To test this, run a test using the loopback adapter - depending on your PC,
the throughput should be very high, but if there is somthing fundamentally
wrong with the UDP stack (or the test app), then it should show up here.
BTW: The app I talked about was written and tested when 100 Mb/s was fast
(NT4 on PII 266) - they are now running on gagabit networks, but i have
never retested the performance.
"JTL" <JTL@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote in message
news:0E6C0319-576A-4612-BCCB-F2B977ACCD7C@xxxxxxxxxxxxxxxx
> "Arkady Frenkel" wrote:
>
>> I mean time parameters just when you change from 1024 to 1025.
>> Arkady
>
> Hello,
>
> Now I have some strange results.
>
> I chose a 3COM PCI gigabit adapter (3c2000). It has a setting for "maximum
> interrupt rate", which is by default 5000. Furthermore, I added a counter
> to
> the sending side so that I now if the sender drops packets (mainly due to
> m's
> suggestions).
>
> When I use packets of size 1024, I receive them at about 250 Mbits /
> second.
> The number might be a bit small, because the testing equipment was not
> state-of-the-art. The packets were received at roughly the same intervals.
>
> When I use packets of size 1025, I receive data at about 40 Mbits / second
> (!), and the packets are received almost exactly at intervals of 0,2 ms.
> That
> is, at exactly the rate set from the 3c2000 driver (5000 per second).
>
> When I use packets of size 2800, I receive data at about 80 Mbits /
> second,
> and the UDP packets are fragmented. The first fragment arrives at time
> zero,
> and the second fragment right after that. The first fragment of the next
> packet arrives 0,2 ms after the first fragment of the first packet. So,
> the
> wire is idle for about 0,18 ms between the packets.
>
> When I set the interrupts / sec to a amount of 10000, the throughput of
> the
> 1024 byte packets is roughly the same. The throughput of 1025 byte packets
> roughly doubles, but the packets no longer arrive at equally clear
> intervals
> than with 5000 interrupts / sec.
>
> None of these tests showed any dropped packets, i.e. the receiver received
> all the packets that were sent from the sender (this I deduce from my
> sending
> program's packet counter that is embedded in the data packets).
>
> It seems to me that 1024 bytes is some magical number inside the UDP stack
> or UDP implementation. What is most peculiar is the case where IP packets
> are
> fragmented.
>
> Any ideas?
>
> Greetings,
>
> Juha
>
.
- References:
- Re: UPD sending perfomance in Gbit Ethernet
- From: Michael K. O'Neill
- Re: UDP sending perfomance in Gbit Ethernet
- From: Arkady Frenkel
- Re: UDP sending perfomance in Gbit Ethernet
- From: Arkady Frenkel
- Re: UPD sending perfomance in Gbit Ethernet
- Prev by Date: having problem conding to deploy application remotely
- Next by Date: Re: Ininterrupted connection
- Previous by thread: Re: UDP sending perfomance in Gbit Ethernet
- Next by thread: Re: UDP sending perfomance in Gbit Ethernet
- Index(es):
Relevant Pages
|
