PCI-Express vs PCI network cards for 1000Mbps

[imaheadcase]

Hi, im going to upgrade eventually to PCI-Express mobo, but right now i have PCI slots only. Im using the mobo lan atm since old network card stopped working, plain old 10/100Mbps.

Now the questions..
According to manufactors PCI-Express network cards are better than normal PCI ones because of bandwidth. Is it really a noticable diffrence? Should i just wait when i upgrade to get a PCI-E one? Does anyone know of a side-by-side review of these claims?

Now im assuming 1000Mbps is just that no matter what one you use, so what our they talking about using all the bandwidth? Or is it just marketing talk they spewing

 

[JackMDS]

Network is Not a Uni Factor system. I.e. the traffic goes somewhere (like to another computer, server, Network related hardware, etc.) and comes from Somewhere (like to another computer, server, Network related hardware, etc.)
As a result the ?Speed? depends on more then one factor (or a piece of hardware).

Current End-User Giga arrangement provides a typical traffic that is significantly less than 1000Mb/sec., and it is Not the NICs fault, so putting an NIC that is alleged to provide more is useless. Unless it is totally in tune with a whole system that can provide more.

Some basics here: Peer to Peer Giga Networks

:sun:

 

[imaheadcase]

Significantly less? I've seen benchmarks showing around 900+ Mb/sec onboard lan. /shrug Think it was on here the nf4 boards.
http://www.anandtech.com/mb/showdoc.aspx?i=2733&p=11

That shows 948Mb/s am i missing something?

Thanks for link

Edit: Ok i read that link, i see the MTU and Jumbo Frames explanation. Thanks for the link again!

 

[imaheadcase]

According to that artical it says we have to wait for using higher MTU settings and Jumbo frames, does that mean you have to get a faster internet connection to raise the MTU to 9000 like they mention? Or is that just OS limit currently?

basicly whats holding us home users back from those speeds? Seems odd its not mainstream for home users yet since all the changes in media now.

 

[BML]

If your internet connection needs a 1000Mbs network card then i want to move to where you live.

 

[imaheadcase]

Mmm i did not say anything about internet connection did i.. Im talking LAN transfers which was pretty obvious.

 

[Madwand1]

I posted the results of some tests I did on this subject in the linked thread. The NVidia NIC referenced could stand-in for a PCIe implementation in those measurements.

http://forums.anandtech.com/messageview...atid=36&threadid=1840646&enterthread=y

The gist is that, assuming your PCI bus is not being heavily used otherwise (e.g. with a PCI disk controller card), for typical usage such as single drive file transfers, PCI gigabit cards can be fine, even without jumbo frames; easily reaching the typical IDE to IDE file transfer rates of around 30 MB/s.

 

[imaheadcase]

Thank you for that info.

 

[cmetz]

Madwand1, the NVidia NIC is not on the PCI Express bus.

PCI-E is interesting because it introduces a new latency in the PCI system: the serialization/deserialization delay. So what you have is more bandwith, but higher latency. Similarly, btw, 1000BaseT has a higher bandwidth than 100BaseT but necessarily has a higher latency because of the DSP required.

 

[Brazen]

Originally posted by: imaheadcase
According to that artical it says we have to wait for using higher MTU settings and Jumbo frames, does that mean you have to get a faster internet connection to raise the MTU to 9000 like they mention? Or is that just OS limit currently?

basicly whats holding us home users back from those speeds? Seems odd its not mainstream for home users yet since all the changes in media now.

Originally posted by: imaheadcase
Mmm i did not say anything about internet connection did i.. Im talking LAN transfers which was pretty obvious.

 

[Fardringle]

Originally posted by: imaheadcase
Significantly less? I've seen benchmarks showing around 900+ Mb/sec onboard lan. /shrug Think it was on here the nf4 boards.
http://www.anandtech.com/mb/showdoc.aspx?i=2733&p=11

That shows 948Mb/s am i missing something?

Those speed tests are for theoretical maximum speeds of a network card (i.e. files transferred from within system memory to system memory on the destination machine assuming ideal network conditions). In reality, regardless of what NIC you use, the speed of file transfers between computers is still going to be limited by other factors, such as the speed of your hard drives. If your drives can only read/write at 25-30MB/s, then no matter what network hardware you have, you'll never be able to transfer files between your computers faster than 25-30MB/s, and the standard PCI bus can handle that bandwidth without difficulty.

If you were transferring files between multiple servers using high speed data systems such as SCSI 360 15K RPM drives in RAID5, then you would use more of the available bandwidth that Gigabit networking offers, but you're still not going to get anywhere near the theoretical limits of Gigabit networking.

Take a large data center (or a few dozen separate servers) all sending and receiving data from super fast drive arrays and you'll stress out a Gigabit connection, but no single computer will ever come close, so worrying about PCI or PCIe adapters really is a moot point. Just pick the least expensive one from a brand you trust.

 

[Madwand1]

Originally posted by: cmetz
Madwand1, the NVidia NIC is not on the PCI Express bus.

Yes, I knew this. I provided the data as representative.

Originally posted by: cmetz
PCI-E is interesting because it introduces a new latency in the PCI system: the serialization/deserialization delay. So what you have is more bandwith, but higher latency. Similarly, btw, 1000BaseT has a higher bandwidth than 100BaseT but necessarily has a higher latency because of the DSP required.

That's interesting. Thanks. Do you have any data on the magnitude?

I also have a Tyan S5151 MB with dual Broadcom NIC's which are said to go through the PCIe bus. So I did some tests with these. Now since the machine also changed, these will not be as directly comparable as my previously linked results, which were all on the same machine. I observed around 109 MB/s receive performance using TTCP, which is close to the 113 MB/s that I got via NVidia, but consistently a bit lower.

Then I observed something even more confusing. The other Broadcom NIC on the same MB, with the same settings, block diagram architecture, etc., identical as far as I can tell, consistently gives around 10 MB/s less throughput. So one port does 109 MB/s, and the other does 99 MB/s. Pretty consistently, despite reconnecting and switching cables, etc. I haven't figured this one out. I guess there might be an underlying architectural difference that's not published in my manual / docs. Or some lost configuration setting somewhere.

In any case, as mentioned a couple of times in this thread already, the differences between 109 MB/s and 99 MB/s and 109 MB/s vs. 113 MB/s are not going to give a significant impact, and often (when using single drive transfers) have no impact whatsoever.

 

[cmetz]

Madwand1, I do not have data on PCI-E systems yet, I haven't gotten any in my lab. Internal R&D budgets are never what I'd like them to be

This is another example of where the NICs' architectures matter. Many NICs have tricks to reduce the number of interrupts, or to interrupt "predictively" based on load. PCI's interrupt latency sucked, and PCI-E by nature should be higher. So a PCI-E NIC will win on available bandwidth, and win again on PCI-E being a switched topology rather than a shared bus, but lose on the packetization and serialization delays. Interrupt latency should be higher, but a well architected NIC can mitigate that.

Now, for extra fun, expect first generation PCI-E parts to be bolt-on/bridged/first attempt kinds of designs. For example, vendor X might build a PCI-E NIC by just putting some PCI-E<->PCI bridge glue into their chip. If I understand PCI-E right, that should be easy to do, and so a lot of vendors will likely do it. And that's unlikely to be good for performance. So many first-gen PCI-E NICs may well be slower than their PCI/PCI-X counterparts.

 

[RebateMonger]

Originally posted by: cmetz
PCI-E is interesting because it introduces a new latency in the PCI system: the serialization/deserialization delay. So what you have is more bandwith, but higher latency.

Ahhh....the old "RAMBUS" problem....

 

[imaheadcase]

Originally posted by: Brazen

Originally posted by: imaheadcase
According to that artical it says we have to wait for using higher MTU settings and Jumbo frames, does that mean you have to get a faster internet connection to raise the MTU to 9000 like they mention? Or is that just OS limit currently?

basicly whats holding us home users back from those speeds? Seems odd its not mainstream for home users yet since all the changes in media now.

Originally posted by: imaheadcase
Mmm i did not say anything about internet connection did i.. Im talking LAN transfers which was pretty obvious.

You are confused because you did not read the link i was talking about..if you are confused don't post. The LAN speed is what i was referring to, again obvious.


Anyways...

@cmetz Did you see in SIsoft sandra they have a comparsion of PCI vs pci-e in the Network benchmark? Pci-e according to that is like 10MB/s faster. Take that for what its work though i suppose in real world. hehe