Network Interface Cards (NICs) 101

[photonboy]

Is the USB2 an issue at 50Mbps or 50MBps?

You said "50Mb" which is just over 6MB which seems pretty low to me since the USB2 controller can manage up to 60MBps so i assumed you meant 50MBps.

USB2 has a 480 Megabit/second maximum theoretical debit, not counting protocol overhead - this translates to 50 Megabytes/second maximum theoretical limit. In practice, it's closer to 40 Megabytes/s.

The use of Gigabit Ethernet at home is easily reached when you have a media server : streaming HD content over RJ45 100 allows one streaming at a time, and the rest of the network (web browsing etc.) slows down a lot. Switching to Gigabit allows for a couple streams and normal traffic takes place as usual. Since a TV recorder is essentially a media server, it counts too.

Uh... thanks but my point was you said "50Mb" which means "fifty Megabits" not "fifty Megabytes".

"b" is bit
"B" is byte

 

[Supermuncher85]

Here I thought I would get an article describing the difference between different 1000/10gb adapters and switches. What a disappointment. I mean common no CPU tests? No bandwidth test?! This was a really lame article....common you can do better.

And at the end you start talking about 5ghz Wifi...like REALLY?! Common guys.... I run a plex server and stream several 1080p streams in my network, and I spent more time measuring throughput in my network than apparently you did.

Check differences between card vendors, and wifi and ethernet not just a generic conclusion of "fiber optics the future" no duh.

 

[2stein57]

Not a very well researched article... please, if you don't know the parts of network card, don't try to name them. The "chip" you named an "EPROM" is a transformer for the signal! And in any case, nobody uses EPROMs anymore, if at all EEPROMs!

 

[g00ey]

Not a very well researched article... please, if you don't know the parts of network card, don't try to name them. The "chip" you named an "EPROM" is a transformer for the signal! And in any case, nobody uses EPROMs anymore, if at all EEPROMs!

If someone had asked me what circuit "4" was I'd say that it would be an optoisolator so it is a little bit of a surprise that transformers now come encapsulated in integrated circuits. It probably does just as good a job as an optical circuit would do.

I'd guess that the EPROM/EEPROM part is integrated with the main circuitry these days. But I still think EPROMs or 'EPROM like' circuits are still used. The BIOS of a motherboard is still some kind of an EPROM/EEPROM, but maybe they call it something else, CMOS perhaps. Also, e.g. android devices use some kind of a chip to store the bootloader and the operating system, I guess they just call it flash memory. I'm not sure whether this flash memory comes built in to the SoC or is fitted on a separate chip on the motherboard of the device. I'd guess the latter but there is a possibility that some flash memory is also located inside the SoC. Also, I'd venture to guess that FPLGA circuits eliminates the need for EPROMS/EPROMS to a certain extent.

 

[GearUp]

I would like to know what others are using to measure performance. I was so disoriented by the last result I reported it as MB/s instead of KB/s which would be high even for hard drives in Performance Monitor. The surprise was from the Mean being the same as the Maximum when also recording TV. That is even without any errors so either the test rides on top of ongoing I/O through the southbridge or it somehow optimizes load to benefit which seems unlikely. Now the result is back down to 28 KB Max total and Mean of 4 KB/s received. I don't think the SpeedTest ever varies so much for download/upload which is 57 Mbps and 6 Mbps.

 

[palladin9479]

I stopped worrying about any of this the second NIC's showed up built into my motherboards.

Most NICs built into motherboards are crap but work good enough for most regular home users. I purchased a $70 NIC to put in my $80 motherboard for my desktop. My mini-server has a $140 NIC in its $80 motherboard.

I only work with Intel NICs at home. Of course there are better brands that Intel, but Intel is the cheapest of the good brands and plenty good enough for me.

Of course some people don't care if they're eating a Big Mac or a home made Black Angus burger.

I'd like to see some benchmarks on this, given I've pushed 117MB/s through an onboard NIC...

What's your CPU usage like when doing 117MB/s? I had an onboard Intel NIC, still Intel, that was around 0.5% CPU while doing 114MB/s. My mini-server is a firewall and was at a messily 5% CPU while doing NAT and stateful firewall at 1.5Gb/s, which was the limit of my integrated Intel NICs on my two desktops.

Once I upgrade my other desktop and put a decent NIC in it, I expect to get a full 2Gb/s(1Gb up+down) through my firewall. During my 1.5Gb/s test, my firewall was only doing a total of 600 interrupts per second load balanced across 4 cores.

To give a comparison to a crappy RealTek NIC, 60,000 interrupts per second all to one core, lots of dropped packets, massive multi-second latencies, and the server was unresponsive during the test. CPU was completely hosed, and because drivers run at kernel level with realtime priority, almost everything stopped working correctly.

The problems your talking about stopped being an issue about five to six years ago. All modern NICs have some form of offload engine with buffering, jumbo frames and most of the standard enterprise stuff. You won't be getting 60K interrupts per second with a modern Realtek NIC. The quality of budge NIC chips has increased dramatically from the old days. A modern CPU can handle many more interrupts then your example assumes, we aren't running single core 200Mhz Pentiums anymore.

The article really should have mentioned line distance and how it relates to speed over copper wiring. 10Gbe with CAT-6 gets it's official distance cut in half, and realistically your not going to want to pure more then 20m. With CAT-6a wiring it's a bit better, 100m officially but that's a perfect run without any signal interference and 10Gbe is very sensitive to EMI.

 

[falchard]

Isn't it mostly useless to have speeds greater than your SSD transfer rate unless you are doing virtual computing. Something tells me 90% of households are not doing virtual computing.

 

[Someone Somewhere]

And I'm over here feeling happy if I can peak 600KB/s down...

Isn't it mostly useless to have speeds greater than your SSD transfer rate unless you are doing virtual computing. Something tells me 90% of households are not doing virtual computing.

I've maxed out (effectively) gigabit ethernet imaging SSDs. Given SATA is 6Gb/s, 10Gb/s gives a lot more headroom than 1Gb/s.

 

[Kewlx25]

I stopped worrying about any of this the second NIC's showed up built into my motherboards.

Most NICs built into motherboards are crap but work good enough for most regular home users. I purchased a $70 NIC to put in my $80 motherboard for my desktop. My mini-server has a $140 NIC in its $80 motherboard.

I only work with Intel NICs at home. Of course there are better brands that Intel, but Intel is the cheapest of the good brands and plenty good enough for me.

Of course some people don't care if they're eating a Big Mac or a home made Black Angus burger.

I'd like to see some benchmarks on this, given I've pushed 117MB/s through an onboard NIC...

What's your CPU usage like when doing 117MB/s? I had an onboard Intel NIC, still Intel, that was around 0.5% CPU while doing 114MB/s. My mini-server is a firewall and was at a messily 5% CPU while doing NAT and stateful firewall at 1.5Gb/s, which was the limit of my integrated Intel NICs on my two desktops.

Once I upgrade my other desktop and put a decent NIC in it, I expect to get a full 2Gb/s(1Gb up+down) through my firewall. During my 1.5Gb/s test, my firewall was only doing a total of 600 interrupts per second load balanced across 4 cores.

To give a comparison to a crappy RealTek NIC, 60,000 interrupts per second all to one core, lots of dropped packets, massive multi-second latencies, and the server was unresponsive during the test. CPU was completely hosed, and because drivers run at kernel level with realtime priority, almost everything stopped working correctly.

The problems your talking about stopped being an issue about five to six years ago. All modern NICs have some form of offload engine with buffering, jumbo frames and most of the standard enterprise stuff. You won't be getting 60K interrupts per second with a modern Realtek NIC. The quality of budge NIC chips has increased dramatically from the old days. A modern CPU can handle many more interrupts then your example assumes, we aren't running single core 200Mhz Pentiums anymore.

The article really should of mentioned line distance and how it relates to speed over copper wiring. 10Gbe with CAT-6 gets it's official distance cut in half, and realistically your not going to want to pure more then 20m. With CAT-6a wiring it's a bit better, 100m officially but that's a perfect run without any signal interference and 10Gbe is very sensitive to EMI.

The issues I mentioned are current real issues. Realtek makes a lot chips with a large range of features and quality. Some are decent but many are crap.

Jumbo frames are crap except for SANs, only work on LANs and have a host of issues they can create. The Internet does not use jumbo frames. LSO has its own set of problems. It breaks TCP pacing, has had security issues in the past, does not play well with all TCP stacks, is mostly non-programmable so don't expect to be able to use custom TCP congestion algorithms, can cause jitter issues because many offload engines attempt to send the segments at full line rate.

No, a modern CPU can't handle more interrupts than what I described in a healthy way. I've seen modern Xeon 8 core CPUs get utterly destroyed by poor interrupt handling. A single interrupt creates multiple context switches, each of which are about 5,000 clock cycles. Even worse is that many of these poorly designed NICs have poor mutli-core support and can cause increased lock contention.

I have a 3 year old laptop at work that consumes about 40% of my quad core i7 just trying to move a messily 80Mb/s. It's a 1Gb NIC, but they use 100Mb ports at work. for non-servers. Interrupts rates are crazy high during these transfers, my computer is almost nonresponsive.

 

[nukemaster]

I have never had network activity slow a computer to any great degree.

Then again I am just doing basic home stuff(all my stuff is over the network on another system so the network get plenty of use in that sense.).

Just a test from my WDC red(3 terabyte model) to my ssd(M4 256) I have an Intel(receiving in this case) NIC in one computer and Realtek(sending) in the other.. I would guess the blip in the middle is the hard drive seeking.

 

[iceman-gr]

You have a mistake on "Anatomy"
Component (4) is NOT an eprom
It's an "LAN ISOLATION TRANSFORMER"

 

[g00ey]

You have a mistake on "Anatomy"
Component (4) is NOT an eprom
It's an "LAN ISOLATION TRANSFORMER"

Indeed it is, for anyone interested in how such a thing looks like inside:

http://p.globalsources.com/IMAGES/PDT/B1123821977/100-1000-Base-T-LAN-Filter-Module.jpg

It probably is just a bunch of small toroidal 1:1 transformers. Here's what an optocoupler/optoisolator circuit looks like:

http://www.azooptics.com/images/news/NewsImage_16276.jpg

good images of such circuits are much harder to find. I'd venture to say that optocouplers provide stronger isolation than transformer circuits, and therefore are better.

 

[Someone Somewhere]

Optoisolators only really work with digital circuits, I believe. Ethernet uses more voltage levels than just 'on' or 'off'.

You also need an energy source on the output side - your ethernet card would be input only (i.e. useless) unless you had an isolation transformer and a bunch more electronics on the port side of the card.

 

[Aris_Mp]

opto-couplers are able to pass some signals that transformers cannot (e.g. DC) while opto-isolators are mostly for AC applications. Also in most of them the channel works only through one way contrary to transformers that can work both ways.

As for WiFi networks you need serious equipment to identify interference and monitor effectively the bands and not toy (hacked) USB dongles. Below are some chanel power measurements from my AC router. In this band currently in my lab I don't have any interference since most users are still at 2.4 GHz which admiteddly is heavily used.

 

[Helix_]

If you are just browsing the web, integrated NIC on motherboard will be fine. Large file transfers do benefit from a decent controller. Case in point : I have a roll your own NAS at home using a low power system (AMD E-350 + 6, 2Tb spinning rust). Integrated NIC had the CPU pegged and SMB transfers @~35Mb/s, added at £20 PCIe intel NIC from ebay and it jumped to ~70Mb/s. NFS is almost wire speed (110Mb/s)

 

[sodaant]

Is the USB2 an issue at 50Mbps or 50MBps?

You said "50Mb" which is just over 6MB which seems pretty low to me since the USB2 controller can manage up to 60MBps so i assumed you meant 50MBps.

USB2 has a 480 Megabit/second maximum theoretical debit, not counting protocol overhead - this translates to 50 Megabytes/second maximum theoretical limit. In practice, it's closer to 40 Megabytes/s.

The use of Gigabit Ethernet at home is easily reached when you have a media server : streaming HD content over RJ45 100 allows one streaming at a time, and the rest of the network (web browsing etc.) slows down a lot. Switching to Gigabit allows for a couple streams and normal traffic takes place as usual. Since a TV recorder is essentially a media server, it counts too.

Maybe if you are streaming a raw blu-ray, but a normal HD stream will not even come close. For example, a 4k netflix stream is only 16-20mbits, so you could quite easily have 3 4k streams, a few music and youtube streams, and still have headroom on a 100mbit connection.

 

[ubercake]

I have never had network activity slow a computer to any great degree.

Then again I am just doing basic home stuff(all my stuff is over the network on another system so the network get plenty of use in that sense.).

Just a test from my WDC red(3 terabyte model) to my ssd(M4 256) I have an Intel(receiving in this case) NIC in one computer and Realtek(sending) in the other.. I would guess the blip in the middle is the hard drive seeking.

I have the same Intel controller on my main PC's motherboard. Does this mean it's better than the usual on-board NIC or is it just a cheaper Intel implementation? Is there any benefit to me getting a discrete NIC card?

 

[nukemaster]

I have not seen a need for anything faster because that has worked fine for me(I am far from an expert on network hardware).

Even the Realtek on the other system seems to hold up fine(I have not checked the cpu usage.).

 

[ubercake]

I guess that's one of those differences I'd like to see in the '... 102' article.

 

[EdJulio]

Thanks for the feedback, everyone. We're building up the Networking category on TH, so these 101's are just the beginning. Reviews, deep-dives and reference stories will soon follow.

 

[turkey3_scratch]

I'm curious how much CPU the encoding and decoding actually uses.

 

[nukemaster]

About 10% on my i5 750.

I do not have exact numbers because other stuff was running eating cpu cycles.

 

[turkey3_scratch]

About 10% on my i5 750.

I do not have exact numbers because other stuff was running eating cpu cycles.

That's actually quite a large amount. Now what I want to see is an article debating whether it is better to simply buy a network card to do this load or a higher frequency CPU.

 

[nukemaster]

I do did not even check to see if the cpu clocks up. It would need a much better test than sending a file across the network while using media center(on its own it does not sit at a steady usage).

To make matters worse, this system is also serving files to my other systems.

I will see if I can get a better idea for you.

 

[palladin9479]

I have never had network activity slow a computer to any great degree.

Then again I am just doing basic home stuff(all my stuff is over the network on another system so the network get plenty of use in that sense.).

Just a test from my WDC red(3 terabyte model) to my ssd(M4 256) I have an Intel(receiving in this case) NIC in one computer and Realtek(sending) in the other.. I would guess the blip in the middle is the hard drive seeking.

It won't have a problem these days. There was a time, in the early GBe days, when the sheer number of packets became an issue for PC's I/O buffers, but that went away once budget chip designs started using buffering and I/O acceleration engines.

For example, I had an AMD E-350 on an Asrock platform running with 8GB of DDR3 memory and hosting a RAID-0 array of 4 x 2TB 7200RPM disks. With Windows Server 2008R2 installed I was able to hit 70~90MB's transfer speed via SMB with Fastcopy and a single giant file. Zero CPU issues on that little dual core. What really hurt it's processor was me logging into the console and trying to do stuff. Eventually replaced it with a hand-me-down A8-7600 and 16GB of memory and another Asrock mini-ITX board with integrated networking. I even went so far as to install WSUS onto it and it's now serving as a local update server for my home.

The only time I had to purchase an add-on NIC was for my home VMWare ESXi lab. I bought a dual GBe port Intel Pro card and only because VMWare has really poor drivers for Realtek NIC's. Most Linux based distros are hit and miss with Realteks, often not all the functions are working right in the open source drivers. On a positive note, having that add-on dual port card enabled me to virtualize my home firewall / gateway device. So now I got my own customized version of CentOS running as my firewall / router / NAT device with two virtual NIC's. Runs great.