Why is my network so slow?

[brown234]

I have two computers hooked up to each other through a hub, and each comupter has a 10mbps nic card in it. When I run the network performance test in Sisoft Sandra it tells me that my network has a max speed of about 300KiloBytes per second. This is my first time setting up a network so i'm not sure if it's supposed to be that slow or not. Is this a normal speed or is something wrong? Thanks for your help.

 

[Vegito]

Due to tcp/ip overhead, you'll probably get around 40-60% of a 10mbps. You wont get a full 10mbps speed, you may also have collisions, etc. even with 2 machines. Are the ISA or PCI nics ? and also whats the speed of your machines.

 

[brown234]

One machine has an isa nic and is a k6-2 400. The other machine is a celeron 700 with a pci nic.

 

[vancur]

I have a few pcs on a network for people to use the internet, but also have some file sharing. Like you, the transfer rate between pcs is very slow. All nics are PCI and all pcs are fast enough to still be fast. I have the TCP/IP protocol. Would enabling IPX speed things up?

 

[spidey07]

300 Kbytes/sec file transfer rate is decent for shared 10Base-T. I would not consider this slow.

To the other question "will IPX speed things up"?
NO.

 

[vancur]

So, is there in fact a way to set up a network where you get transfer speeds as fast as those when downloading from the internet? If not, it doesn't make much sense to me... Shed some light on the ignorant.

 

[CTR]

Spidey has extremely low standards. That explains a lot about him... Maybe he can provide us with a packet analysis comparing IPX and TCP/IP. That's what he really wants to do. So I'm going to say:

"WTF Spidey07? IPX _is_ faster than TCP/IP. You ever hear of packet burst?"

vancur: The more money and time you spend, the faster and more reliable your network will be. :Q

brown234 should check his duplex and speed settings, and watch for any collisions on his ether stats.

 

[spidey07]

<< So, is there in fact a way to set up a network where you get transfer speeds as fast as those when downloading from the internet >>



Wow, that will require some preaching:

Network speeds are set in stone. For instance 10Base-T has an effective capacity of 10 Megabits/sec (even though the baud rate is really 12.5 Mbit). Now does this mean you will realize 1250 Kbytes/sec transfer rates between two computers on a 10Base-T net. Heck no. But like most network engineering the answer is &quot;It depends&quot;. First off it is difficult to utilize much more that 40% of the capacity of shared ethernet. This is simply a limitation of the layer 2 protocol CSMACD. There are all sorts of timers involved dealing with inter-frame gap and collisions that make this %40 rule a nice one of thumb (Now if somebody comes back and says &quot;well I've seen %80 utilization&quot; then fine, but 40% is the accepted generalization)

Moving up the OSI stack since we covered limits of Layer2 we are at the IP layer. No real limit here but some more overhead.

Layer4 - TCP. Now this has the biggest impact on transfer rates (discounting HD speed, file caching, cpu, NIC, etc). TCP does all the flow control from sender to receiver and is responsible for windowing. More payload overhead to deal with.

Now you understand limits of ethernet and you understand several required layers of overhead. First off when downloading from the internet you are using well know applications such as HTTP or FTP. These applications have been around for decades and are quite effecient at what they do. To be put another way...windows networking (SMB) flat out blows. There is too much overhead in the form of DATA over SMB over NETBIOS over TCP over IP, etc. Very poor performance not to mention windows to not dynamically maximize their TCP window. This is why you should modify your RWIN paramater in the registry. Add all this up and maybe, just maybe i've rambled on long enough to show you why you'll never get much more than 300-500 kbytes/sec of PAYLOAD transfer speed from shared ethernet on any given TCP connection. Especially with windows file sharing.

Want to get faster? Use a download accelerator like, err...&quot;download accelerator&quot;. These over come the limitations of using only a single TCP connection by using multiple ones therefore maximizing the use of your bandwidth. During a file transfer a significant proportion of time is spent waiting for acknlodgements from the recevier. If you have more connections then chances are while the server is waiting from an ACK from you you are still receiving more data on another TCP connection, so instead of just waiting while ACKing your actually doing something.

And if you want even more performance then by all means use a 100 Base-T switch with full duplex. 100 megabits payload and you eliminate and backoff timers used for collisions because there will be no collisions. Expect to get payload transfer rates of 3000-5000 Kbytes/sec

Way too much information, but I'm ramblin.

spidey

ps - I don't even have a clue as to how SANDRA measures network performance. That number might be totally bogus.

 

[spidey07]

CTR - yeah, yeah, yeah. I hear ya. Does microsoft properly implement packet burst?

What, you never heard of LARGE TCP windows? That blows packet burst out the door.

CTR - don't make me do protocol analysis. Why when I was young I had to do it with a hex editor no fancy sniffers. my head hurts already. Where's the beer and napkins?

 

[CTR]

Point taken!

Microsoft thinks packet burst is when the packets collide and leave sticky residue on your hub.

Novell did it really well, though.

 

[Rolkin]

Why do you have a hub if your using 2 computers why not just make a crossover cable, and lose the hub. I have a 10/100 hub with 10/100 nic cards but losing any extra connections would...I don't know just make me feel cleaner about the connection.

 

[CTR]

That is certainly a valid point, since the problem could very well be the existing cables. Why not start at OSI Layer 1 and work your way up -- a tried and true troubleshooting system.

1 Cables/Hubs
2 Switches/Bridges (there are none in this case)
3 Routers/NICs
4 Operating System

 

[queyan]

I'm using a BNC cable between my 2 computers. Should I change to a crossover cable to get better performance (online gaming, internet, file sharing, etc.)? Add a hub maybe? I'm currently set up to share a wireless DSL connection using Analogx proxy program.

Rod

 

[brown234]

I'm using a hub because i'm about to add another computer to the network. When I was using a crossover between the two computers I was getting the same speed results. What is a good program to use to measure my network performance? Thanks.

 

[Emory7]

Sandra says my throughput is 1114kB/s. I have three computers connected via a hub 1200ft apart.
Cards are 10/100, 5 port generic hub I paid 45$ for, dunno name.

 

[Ringer]

I am running two comps with a crossover setup. I share a printer and broadband cable. I use sygate home network software and have no problems. I also have sissoft2000 pro and and am getting over 1000 mark on each comp. There is really no need for a hub or switch when 2 comps are used. I use two nics in my server comp, one for the broadband modem and one for the client comp. Play multiplayer all the time online and in a lan mode, outstanding pings. I really dont know too much about networking, but this is a simple as it gets. I highly suggest you look into sygate home networking software and try the crossover cable method. Hey, its free...



Text

 

[spidey07]

OK, just got home and did a sniffer trace (thank god no HEX) of the SANDRA network benchmark. I'm actually impressed, this benchmark calculates theorectical network performance with a very real methodology. It does it in an interesting way though. All my blabering in the above post was concerned about overhead and TCP flow control...all of which can have a big impact on performance. SANDRA's benchmark has nothing to do with flow control or even TCP for that matter or even usable performance. It simply sends out a ping of 65000 bytes and calculates the response time. Using simple arithmatic you can use this to calculate network performance (discounting everything that has to do with &quot;real world&quot; network performance as I mentioned in my previous post however)

Now the big kicker is the portion of the network benchmark that says &quot;HOST&quot;. This is the other host that SANDRA pings. If you don't have another PC on your LAN or you have MS networking somehow screwed up (it looks like SANDRA picks a host from the MS browse list or is it ARP cache?, that's why I got a local host) I can see why your number would look bad.

On my very simple home LANs I consistently get 1114 KB/sec on Sandra 2000 7.6.49. 1114 KBytes * 8 bits = 8912 Kbits = 89.14% utilization on ethernet. All in all I would expect this kind of benchmark to perform this way. MAN IS THIS MISLEADING IN TERMS OF REAL WORLD PERFORMANCE.

But to answer the original question:
IF THE HOST YOU ARE TESTING WITH IS A LOCAL HOST (no router, same IP network, same hub) THEN 300 KBYTES/SEC IS INDEED VERY SLOW USING SUCH A SYNTHETIC BENCHMARK SUCH AS SANDRA. Go with CTR's advice and make sure cabling is good and you aren't picking up any errors on your NICs. Also don't use crap nics such as linksys. Intel all the way baby. I'm gonna sound surly here but if you use crappy nics then you have no reason to gripe about performance. Same with everything in life/technology. You guys know what happens when you buy el'cheapo condoms right? Bad mojo.

BUT IF YOU ARE NOT TESTING WITH A LOCAL HOST THEN YOUR RESULTS COULD VARY GREATLY, at that point you are taking into consideration router processing and overall network latency.

Hope this helps.
Spidey

PS - LET ME REPEAT IN BOLD LETTERS...AFTER LOOKING AT THIS TRACE THE SANDRA NETWORK BENCHMARK HAS NOTHING AT ALL TO DO WITH REAL WORLd PERFORMANCE OR THROUPUT.

 

[CTR]

Spidey you are the master of CYA.