Speed & CPU Usage with D-Link 650

< >...



Yeah. To say the least D-Link has a good design, except they screwed the
pooch in R&D on this one. Modification of the antenna is *very* easy in this
model. (believe me, took me all of 10 minutes to add my own antenna)
However, the internal antenna is vertically polarized. Can anyone guess what
that means? heh. Learn to type sideways comes to mind. :) However they are
powerful, 14-15dBi, iirc.

Driver optimization for windows is horse piss on these things too. My XP
laptop killed one of these cards because it automatically updates firmware
when you install the drivers. *VERY DUMB D-LINK!*

However, my other UNIX laptop, does not have cardbus, and these things work
flawlessly in there. (They're not keyed for cardbus only slots) Work
perfectly, however, again, since the drivers are written by someone from
FreeBSD, they're not quite optimized either, but better than the windows
drivers. Go figure.

Also, i don't know if this has been disscussed or not, (stupid unix news
program), but if you're running WEP also in your config, EVERY packet has to
be decrypted, which would drive the load on your machine through the roof.
If you have any 2.4Ghz cordless phones, move them, or just unplug them
temporarily to see if that improves your signal. My old boss said his 2.4Ghz
baby monitor was being interfered with with wireless networking, so it could
go both ways.

Hope this helps ya.

--
Alex Hartman - goo @goobe.net
PGP Key fingerprint = 26 41 19 56 19 81 E2 BC  EE C8 1D F4 DB B8 ED B8
"Watch out for that bus!"

<snip>
<snip>
Yeah - I think we are coming to the conclusion that 802.11b card drivers
suck.
I have problems on my iPAQ 3850 when trying to run encryption - the device
driver keeps dying. I suspect that this may be due to lack of processor
'bottle' despite having the latest Strongarm processor.
We have proved that the PCs can shift data at a high speed if there is a
decent network underneath.
Therefore any loss in throughput must be down to the network components.
The suggestion that they are 'Winmodem' style cards could be accurate.
Even so, a 1GHz processor should be able to shift data at 11MBits/sec as
long as there is a reasonable size I/O buffer.
Perhaps there is something buried in the protocol or in the way the cards
turn round from transmit to receive that limits the real throughput of the
system.
The AP must also be doing background housekeeping such as transmitting to
let clients know it is there, and there must be some overhead in the AP
receiving a packet, buffering it, and then retransmitting it over the UTP.
The problem could even be a lack of buffer space in the AP, causing it to
flow control.
It would be interesting to compare the throughput of two DWL-650s in
'ad-hoc' mode which would remove the AP from the equation, and also remove
the retransmission within the AP as a source of delay.
If I get a PCI card to take PCMCIA cards I might try this - or just try
copying over wireless to my IBM Microdrive, although I am not sure what the
maximum transfer speed for this setup is and don't have a suitable Ethernet
PCMCIA card for the iPAQ to cross check the results.

If anybody has two PCs which are set up to talk in 'ad-hoc' mode they might
like to supply some results?

"David W.E. Roberts" <nos @talk21.com> wrote in message

On the same computers I can copy files with 11MBytes/s over 100mbit wire.
And that doesn't even take as much CPU as the 500-600KBytes/s wireless
transfer does.


Yes, true. And if Alex suggestion that 11mbit was the full utilizing with
full duplex, then we could compare that with marketing a normal 10/100
switch as a 20/200 switch.

/J

You know Alex, I hadn't really thought about that but ofcourse it is.
One antenna and one channel ipso facto...  Perfect explaination,
elegant as well.

I really should have thought of that.  Lets see then...

Tnx
-m-

I don't think this is how it works; half duplex means only one station can
send but each station can send at 11MBits/sec when it has the link, and if
one is sending large chunks of data and the other is just sending short
acknowledgements then there will not be a 50/50 split on the link. You will
not get the full continuous 11Mbits/sec in one direction, but you should be
able to get close with the right setup.
To really test the maximum throughput you need two tiny programs, one
sending large blocks of data from memory, and the other receiving the data,
acknowledging, then throwing it away.
You also need nothing else getting in their way when they want CPU and I/O -
a lean mean operating system and no other processes. More likely to get this
with Unix style O/S than Windows.
If you are copying files then the virus checker will be taking a keen
interest as well.

<snip>

Well, you can only send data at 100MBits per second.
However, if you are runing full duplex then you can receive data at 100MBits
per second at the same time.

Full duplex means that you can get your protocol acknowledgements back at
the same time you are sending more data. With half duplex you have to wait
for an 'ack' once you have sent your window size of data. So potetially you
could send contiuously at the maximum 100MBits per second with full duplex.

I assume this only works with twisted pair implementations where you have
send and receive data wires. The old CSMA/CD Ethernet was just a bus - only
one station could transmit at once so full duplex was not possible.

I am running at half duplex - didn't see the point at the time when I bought
my Netgear hub to pay the extra for the full duplex. Nothing I do will fail
because I cannot utilise the full 100Mbits/sec in both directions
simultaneously, and the 64k ISDN link is not going to notice the difference.

Cheers

Dave R

That is what is being reported at www.practicallynetworked.com.  I
think that checkbox must be something to do with the wired connection
then, though I can't imagine why that would be necessary.



You are welcome, sir.
-m-

Interesting. So what you say is that in the world of wires we got
20/200mbit ethernet, not 10/100 as I was told before ??? ;-) I buy what
you say, but I think it is false advertising if thats the case.




This is by doing a normal copy using drag and drop in windows or by ftp to
my local server. I get 60-80% just transfering data at maximum wireless
speed and another 60-80% decompressing the movie which sums to framedrops.

I got 3 different ideas why the *HUGE* CPU-load:
1. The DWL-650 works like a winmodem. The CPU does all calculations and
the DWL-650 is only the radio RX/TX.
speculation: Why does it then work at all on a P-120 laptop?

2. The Cardbus utilizes an HUGE amount of interrupts compared to
PCI-communication.
speculation: I dont know much about cardbus, but I guess its more of a
bottleneck than PCI?

3. The driver is very poorly written/designed.
speculation: Linux seems to utilize less CPU. But thats comparing
apples/oranges.

Im curious about how other 803.11 PC-Cards do?


I was running on power when doing my tests, and I could n't tell any
difference when running on battery.

Thanx
/J

By nature, 802.11b is half-duplex. unless you buy some expensive breezecom
equipment, 6mbit/s is where she tops off.

So yes, that is normal.

Depends on the laptop, depends on the compression of the movie you're trying
to watch. If it's divx (i do the same thing here at home with an almost
identical setup to yours.) then yes, it's going to eat the CPU for lunch.
Not to mention it's a laptop, if you have it unplugged, can anyone say
SPEEDSTEP? I do this with DVD backups, and i get occasional framedrops if
someone walks in front of the machine (which is sitting ontop the stereo,
and the AP is ~100ft up and to the left) otherwise it's fine.

Speedstep man, Everyone ignores it til it eats the battery.

--
Alex Hartman - goo @goobe.net
PGP Key fingerprint = 26 41 19 56 19 81 E2 BC  EE C8 1D F4 DB B8 ED B8
"Watch out for that bus!"

The rest of the test is embedded.  The Laptop is a dual boot system.
W-98 and Linux.  I got up this morning and tested the W-98 performance
using exactly the same test.  I had not moved the laptop access point
or changed the test file.

Linux was....

W-98 was ...

  ~500 KBytes/Sec sustained with 90% CPU Load.

The Windoze 98 CPU load is (.9*180) which is about 160 MHz and still
way less than 60 percent of an 1100 MHz processor, more so because the
newer processors can have so many more instructions in the pipeline at
any given time.

The problem with Windoze performance seems to be windoze and it
appears to be getting worse with each iteration of their OS.

--
"Microsoft; 98% market share can't be wrong, even when it is wrong."

Damn sure it is not running at 10Mbit if I'm getting 13.6Mbit from the link.
I don't expect a filestore hungry application which is opening and closing
loads of files, creating at the far end, scanning directory structures etc.
to achieve flat out 100MBit throughput on a LAN.
Remeber also that 100MBits is the speed of the individual bits of the data.
This doesn't mean that all LAN cards have the power to maintain this speed
in a contiuous burst, even if the PC can push/pull this amount of data to
and from disc. A lot depends on how the driver is written, how much buffer
there is at each end, packet size, window size, etc. Also allow for
collisions with other users of Ethernet.

I tried to set High Speed Network, but I could not see any difference :(
I also tried to enable wep and I get the same speed and CPU-usage with wep
activated.


I will try Linux to see if there is any difference.


True, but im doing the same while running on wire (100mbit) and I get
different readings. Its interesting to see though that you get different
results using a P120 running Linux.

Thanks
/J

The other replies made me do it...

Linux 2.4.x on P-180 MHz Cyrix Laptop w/32 M and 2 GB IDE, not nearly
the machines you folks are discussing.

Same wireless hardware exactly.

10 MBytes FTP between to Linux boxes, the second box is Linux 2.4.x on
Cyrix P-133 w/32 M and 2 GB IDE.  I am pushing the file from the
laptop to the dial out server thru the AP.

~600 KBytes/Sec sustained with 28% CPU Load.

Data transfers are about the same speed, CPU load is (normalized)
about one tenth as high, if you are seeing 60% loads on a 1.1GHz CPU
and I am seeing 28% loads on a P-180 [(1100*.6)/(180*.28)].

I suspect checking that "high speed network" box will help the
transfer rates. Nothing will help the CPU usage...  except spending
less money on your OS.

:P

-m-

Really?  Are you *sure* it's running at 100 MBit?

-m-

Have you set the checkbox in the AP which says "High Speed Network" or
something  like that.  It may not even refer to the wireless network.
I ask only because I observed it was there.  I never read the
documentation durning setup so I really don't know whether it applies
to the wired or wireless portions of the network.  I'd expect it to
apply to the wireless portion *BECAUSE* the hardware can figure out
which the wired network is without help from the OS.  I did not check
mine and I observe about the same speeds you observe.

That would likely be an issue with Windows or the driver software.  On
an old P-120 running Linux, I see no measurable loading running
500KBytes/sec file transfers.

No, generally any processor running over 100 MHz has so much more
bandwidth available to it internally that it *should* be capable of
completely swamping a 100 MBit/sec link with less CPU loading than you
mention, if that is all that it is doing.

Windows does use a *bunch* of CPU to keep the screen up to date, when
there is something to update.  Are you running an app that requires a
lot of up dating of the display to monitor the network?  The Microsux
performance monitors (indeed top on *nix systems as well) require a
lot of CPU to monitor system processes because they have to execute
between each task swap, hence they execute *WAY* more than any other
process.  Top is good to tell you how much CPU it is using, not sure
about M$ software.

-m-

Wish I could help you more, Jonas, but I don't know enough about the CPU
participation in wireless communications.  I wouldn't expect it to be very
CPU intensive.  I have an old 266 MHz IBM ThinkPad I use with wireless and I
do get acceptable throughput, but I'm not trying to watch movies on it.  If
you can buy 802.11a gear with satisfaction guarantee, it might be worth a
try!

Don W.

"David W.E. Roberts" <nos @talk21.com> wrote in message

Thank you David.

My Dell Inspiron 8100 P1.1Ghz P3 WinXP gives about 10% CPU per 10mbit
which means I get about 60% CPU when having 60mbit speed over the wire
(100mbit).

When doing the same over wireless I get 60-90% CPU with about 4-5mbit
(550KB/s).

There is one big difference. The ethernet wire connection is based on
intel chipset and is a mini-PCI device and the 650DWL being a PC Card.

Regards
/J

Well, when I connect it through wire (100mbit) I can have even up to
11MByte/s with lower CPU-usage. And that doing the same thing. So I think
there is something else. For example, does the CPU handle all
"DSP"-calculations, and the card only being a media? Does that differ
between manufacturors/chipsets?


Yes, no problem at all. No framedrops and I cant even tell if im watching
from my server or locally from harddrive when looking at taskmanager.

But how sensitive is 802.11b? Is it possible to establish an environment
at a place like a home and recieve 11mbit?


Yes, it is a DWL-650.

Regards
/J

"Normal" depends on what your're doing with the data.  Just talking to the
network isn't taking 100% of your CPU time, but storing the data on disk may
take some time or decompressing a movie is extremely CPU intensive!



The reason the movies use 70-100% of the CPU is the task of decompression.
If you connect the laptop to the server via wire, are you able to watch your
movies without framedrops?  If a wired connection works ok, then sell your
802.11b equipment and buy 802.11a.

You're right about "overhead".  Without any lost packets the overhead is
less than 50%, but radio isn't perfect and if any packets need to be
retransmitted, then throughput suffers.

I'm assuming your D-Link 650 is a DWL-650 and not a DSB-650 (USB version.
There can be a bottleneck with USB -- particularly if you're sharing the bus
with other devices.)

Don W.

Jonas,
just done some tests - interrupted by BSOD when I swapped network cards
whilst the LAN analyzer was running.
Setup:
WinME 1GHZ Athlon with 100Mb Ethernet to NetGear 100Mb hub.
DWL-1000AP in the next room.
Win2k 700MHz Pentium 3 with DWL-650.
Monitoring using LanScan, CPU via Task Manager.
Generating traffic by copying about 900Mb of MP3 files using Windows
Explorer from ME to 2k.
Results:
LanScan shows about 350,000 bytes per second; roughly 2.8Mbits/sec
This matches what the throughput in the status window shows for the DWL-650.
Link and Signal around 50-60%
CPU around 60%
Cross Check:
Repeated with 3COM PCMCIA card running at 100Mb/sec.
LanScan shows about 1.6Million bytes/sec (plus 0.1M coming back) so about
13.6Mbits/sec.

Conclusions:
You aren't getting the full bandwidth; a shame I don't have 10 Mb cards for
a comparable test - this would show roughly what overhead you get on a
standard Ethernet system. If I'm bored I may force both cards into 10Mb/sec
and try again, but tea is ready so I'm off.

Cheers

Dave R

"David W.E. Roberts" <nos @talk21.com> wrote in message

Im doing it very simple. Running the taskmanager (winNT, 2K or XP) in the
background and using a utility called DUmeter v3.0 to measure bandwidth.

Also Im downloading through ftp from my local server which gives about the
same result as DUmeter.

DUmeter is 30-day shareware at www.dumeter.com

Please tell me the CPU-Usage when transfering a file over the wireless
net. If you got Win9X/ME then try a utility called "netstat live" at:
http://www.analogx.com/content s/download/network/nsl.htm

Thank you
/J

Jonas,
how are you measuring your throughput?
I have a D-Link Wireless Kit and I could repeat your tests just to see how
mine does.
HTH
Dave R

So there is 50% overhead with the wireless media itself?

There were no encryption, wep or anything like that involved, only "pure"
tcp/ip.

What is "normal" without any datacompression?


Well, unfortunately there is a problem.

I'd like to watch movies I have on my local server. They need about
200-300KB/s so that shouldn't mean any problem. Though these movies
themself uses about 70-100% CPU so I get framedrops from not enough
CPU-power.

btw. my internet connection is 2mbit

Thanx
/J

4-5 Mb/s data transfer rate is not bad, and depending on how you measure
that, you might even see 6 Mb/s, but no more than that.

Your CPU usage has a lot to do with how you're handling the transferred
data, the protocol you use to transfer the data and any
encryption/decryption/compress­ion/decompression used.

If your setup is for wireless internet connectivity, neither that speed nor
the CPU usage should be a problem.  Most residential broadband connections
are less than 1 Mb/s.

Don W.