Date: Thu, 29 Jan 2004 20:33:48 +0100
From: Philip Nienhuis <[EMAIL PROTECTED]>
Subject: Re: [LIB] Networking problems

Raymond wrote:
> >From: "Matt Hanson" <[EMAIL PROTECTED]>
> >Subject: Re: [LIB] Networking problems
<snip>
> >Also, since I tested xfer rates both between Libbys and beteen each of
> >them and the desktop with the same results, I guess components are not the
> >problem.  I guess it's just a limitation of the CPU, and what supports and
> >augments it.
> 
> Hi Matt,
> 
> There is a logical explanation for all this and I'm almost certain the
> problem here is CPU power. There is a big difference between dumping
> digital videos through firewire at almost 30Mbps and only managing 5Mbps
> through a network card. Firstly, firewire cards (and, to a lesser extent,
> USB2.0 devices) have onboard controllers. In my card (and I assume almost
> all firewire cards), the firewire stream goes straight to system memory via
> the DMA (direct memory access, allows devices to read/write to/from main
> memory independant of the system processor) controller, bypassing the CPU
> altogether. All the CPU has to do is grab the data out of memory and shove
> it out to disk (the video is uncompressed and, as far as I'm aware, comes
> across the firewire cable in a format that is almost like uncompressed AVI
> anyway). In fact, if the video is already interleaved correctly, all the
> CPU needs to do is monitor the DMA controller until it says it has a block
> of data then tell the IDE controller (also on DMA) to grab data from that

Are you sure about DMA in the Libretto?
AFAIK the Libretto 30-110 IDE (which is attached tru a plain vanilla ISA
bus)
simply has no DMA support. PIO mode 4 is the best I could ever get on my
Lib110 in Win2K, OS/2 and Linux.

<snip>
: 
> In contrast, a network card presents the CPU with network packets which
> must be processed, error checked, resent if corrupted and so-on as
> specified by the various protocols involved (TCP/IP and SMB if you're using
> Windows filesharing over IP for instance). The CPU needs to do a lot of
> work before it can unwrap the data encapsulated in a network transmission
> and I can well believe that the poor L70 or even the L110 will croak once
> you get around the 6Mbps mark.

NIC brands do make a big difference, as some take 60-80 % CPU time,
especially in 100Mbps mode, while other brands do a lot by themselves
and leave more CPU time for other processes. Cheap NIC cards rarely
perform as well as e.g. 3Com, Xircom, ....
It is really a matter of value for money here.

Just for reference, I checked the transfer speed here with 3 LAN
PC-Cards: (Lib110 <->
1.6+ Ghz AthlonXP all Win2K), copying 79 MB with plain NETBEUI/Netbios
(not over TCP/IP, as that is 5-10 % slower due to TCP overhead):

- Xircom CEM33 (10 Mbps / full duplex): 2 min 18 sec
- Conceptronic CON100TCL (100 Mbps / full duplex): 0 min 47 sec
- IBM 10/100 Etherjet (100 Mbps / full duplex): 1 min 02 sec
(oh and yes, I got a switch in between),

so the fastest effective rate I got is in the order of 1.4 MByte/second.
I'm afraid that's
as fast as one can get on a L110, probably slower on a L100 (lower CPU
speed), not to mention L70.

BTW It's remarkable that the 32bit Cardbus IBM Etherjet (actually
rebranded Xircom) is actually slower than the 16bit Conceptronic.

Philip



**************************************************************
http://libretto.basiclink.com - Libretto mailing list
http://www.silverace.com/libretto/ - Archives

                 -------TO UNSUBSCRIBE-------
Reply to any of the list messages. The reply mail should be
addressed to: [EMAIL PROTECTED] - Then replace any text
on the message's subject line: cmd:unsubscribe
              --------TO UNSUBSCRIBE DIGEST------
Do above but with this on subject line: cmd:unsubscribe digest
**************************************************************

Reply via email to