N-BASE T- anyone using it?

I have a few devices with 2.5Gbit/s interfaces. It does seem to be almost mainstream now, even in very low-cost Chinese PCs. For example, I have a sub £200 N100 PC which came with two 2.5Gbit/s ports. It works fine over CAT5e cable and is rated for 100m connections at that speed. There are cheap USB3 2.5Gbit/s dongles which also work fine. Its nice to have, but as for needing it - well one day perhaps? Some WiFi access points have 2.5Gbit/s interfaces as otherwise the wired ethernet would be the bottleneck in throughput. Connections to a printer or NAS device might benefit. It would almost certainly speed up backups. There are also 5 and 10Gbit/s over twisted pair interfaces, but they require better cable such as CAT6A for the full 100m performance. CAT5e does work at 10Gbit/s over shorter cable lengths. Power consumption is a lot higher for these speeds. John

My mate's FTTH is 900mpbs and is currently under £25/month.

It certainly does, My laptop (over thunderbolt) and server (PCIe NIC) both have 2.5Gbe

It is the latter that would probably drive the requirement - say fast access to a NAS. Even a run of the mill hard drive will more than saturate gig ethernet, and a NAS with (say) RAID 10 (mirrored and striped) could make good use of 2.5G ethernet.

(mine has 4 bonded gig ports - so any one machine is limited to gig speeds, but it can service more than one with full gig speeds to each)

Not personally

Or slightly posher switches with some special function ports, will take a variety of plug in interface devices. Can be handy for the backbone between switches.

They already are - but quite often the internet servers don't keep up with them :-)

The spread of 2.5GbE ports can be blamed on RealTek. On retail motherboards, it is far and away the favoured cheap manufacturer of NICs. They used to put

If all the LAN machines had it, you could search for a 4 port switch, to allow faster communication between machines (LAN file transfer).

Other (home user) speeds available, are 5GbE and 10GbE. At the 10GbE level, the cards for that were Enterprise versions, and expensive. The Aquantic AQC107 was the first chip to attempt to do a lower cost version for punters. The company may have changed hands a few times. But the cards are still available at retail. The cheapest way to enter that domain, is put a card in two desktop machines, and not even connect to a router or to a switch. Just a private high speed connection between two powerful machines. That's the cheapest way to start. Some of the NIC boxes, even come with a "scrap of cable in the box", for a connection to the other machine.

To do a switch with (5) 10GbE ports on it, now that would be expensive. That would be around five hundred, and you also have the issue of the switching chip needed to support gobs of bandwidth.

When it comes to "routers", there is the speed of the WAN side, but there is also the routing capacity inside the box. If the CPU inside the router is not fast enough, it cannot "route at wire speed". For example, the routing in the room here right now, has a rating of

In the client PC OS, the OS "handles interrupts". Windows has an "interrupt limiter" which keeps the OS working in cases where a card artificially keeps the interrupt signal asserted permanently. It's a kind of safety valve, if for example, the OS is to print some warning on the screen. If there was no interrupt limiter, the OS could stay at driver level permanently. And just appear frozen (when it's not actually frozen, it's just very very busy).

So, OK, we have an interrupt limit. Has the interrupt limit been made proportional to the power of the CPU ? At one time, it might have been

This also means, eventually, the OS has an "upper limit" on traffic level. Sure, you bought a 2.5GbE NIC, but is the OS rated for it ?

On Win2K, I did some analysis, and 40MB/sec out of 112MB/sec is the most I could get. Win2K was not ready for the 1 GbE hardware that eventually showed up. And interrupt consolidation might not have existed at that time. I have not hit a limit since then. WinXP could do at least 112MB/sec.

The absolutely worst/rubbish NIC, used to deliver *five* interrupts per one packet. It looked like an attempt at "zero copy", and placing packet header in one place via DMA, packet body via DMA somewhere else. The result was, the GbE NIC involved, could only do about 70MB/sec just due to hitting the interrupt limiter. This might well have been a RealTek NIC. By extrapolation, I figured my ~3GHz processor would have to run at 4GHz, just to service the stupid NIC. Other NICs are not like that. And consolidation is one way you get more miles per gallon.

So while as hobbyists "the sky is the limit", there are a few rough edges under the hood. There have been some used 100GbE cards (might fit into a x16 PCIe slot or the like) on Ebay, but I haven't seen any careful testing of whether available OSes could make productive usage for transfers within the computer room. Or even, if there is a driver that works for such a thing.

Generally, nothing scary will happen with a 2.5GbE NIC. It's not enough of an evolution, to particularly care about rough edges. But before I buy a 10GbE card (a gigabyte per second between two machines having RAM disks), I would need to find a test result which tells us how much headroom the modern OSes have for this. And within the routers, we would also need to discover what the WAN-->LAN routing performance is like.

On the topic of fibre to the home, I find it strange that any business can afford to build infrastructure (routers in the network), fast enough to deal with all these whizzy rates.

Google is apparently in possession of PHY electronics (E/O) capable of 20Gbit/sec rates on a per-home basis. But nobody believes they'll be stupid enough to actually run at that rate, and if those are deployed, the rate will be "somewhat better than 1Gbit/sec". That's a test in Kansas City, Kansas.

The fastest rate here, is an (inconvenient) 3Gbit/sec, and the number of buildings where that is available, is not very great. The "boxen" at the corner for that, must be a bit more expensive, and then you need customers with bags of money to pay the monthly for it. The bleeding edge people who test stuff like that, usually find rough edges and "limits" not entirely consistent with the service.

If on a 1 Gbit fibre here, you got 800 or 900 on upload direction, you would not be completely surprised by that. I haven't seen the bench results for 3Gbit, as it's going to be difficult for a home user to properly provision for test of that. People were buying replacement modules for some 1.5GbE service, to try to get the full rate from it. Places like DSLReports, you would find technicians making suggestions for home users, on how to kit up for a full rate setup. But generally speaking, if you volunteer for >1Gbit/sec, you're like a sheep to be shorn (paying too much).

Surprisingly, you can get better rates from some Internet servers. When they configure servers, they cap a connection. With archive.org, that might have been 345KB/sec. But on the other hand, the server operator also allows way more than 2-3 connections per home. And if you use aria2c or the like to open multiple connections and do byte-range downloads, you can do a better job of filling your link. I don't have fast enough internetz here to test that ("cheap" is my middle name).

Summary: 2.5GbE. Just plug it in and use it. "Don't you worry" :-)

Paul

If you're going to do point to point, better go with SFP+/QSFP rather than RJ45. Ex-server cards are available cheap: I bought some 40GbE QSFP NICs for £40 each. Then use a SFP+/QSFP direct attach cable to connect them.

It's about £150 now for a 4x 10GbE + 1x 1GbE switch:

I did get ~35-37Gbit/s in iperf, so Linux could handle it. I agree apps can be a bottleneck.

It can be done for Windows too:

B4RN will give you 10Gbps in the sticks of rural Lancashire for £150/month:

Theo