12

This is another one of those hypothetical questions. I've been trying to figure out whether or not having a faster 'segment' of a network in between host A and host B will translate into a faster flow rate or lower latency between them. Let me show you the physical links in the network between computer A and computer B:

host A (1000Base-T NIC) -> copper 1000Base-T link -> 1G copper switch -> 
[SFP module] -> a short 10G/40G/100G fibre run -> [SFP module] ->
1G copper switch -> copper 1000Base-T link -> host B (1000Base-T NIC)

In short, there is a 1G link from host A to the first switch, which has an SFP module in it connected to a short 10G/40G/100G (doesn't really matter, just faster than 1G) fibre run, which connects to another SFP module in another 1G copper switch, which is connected via 1G copper to host B.

Does traffic flow faster between the two hosts because of the fibre run in the middle? Or would the flow rate and latency be the same if the section between the two switches was the same speed as the rest of the network?

It would make sense for latency to be lower between host A and host B, but the ingress and egress rate of the NICs would limit the flow rate, correct? If this is so, does it make sense to connect 'core' switches and routers together with faster links?

13

Neither, really. Replacing a copper link with a fiber link might lower latency a tiny bit (assuming an uncongested link), but what you really get when you replace a "core" link with a higher bandwidth link is less possible congestion. In your example scenerio, it doesn't matter, because there is only one device on each end. In a live network, though, moving from 1g to 10g core links will ease congestion problems within the core of the network.

Now, as a side affect, you may get lower latency and better traffic flow, but that is purely due to easing congestion so the routers/switches aren't overloaded and dropping/queueing traffic.

| improve this answer | |
  • So increasing the speed of the links in the core of the network increases capacity, not speed. – Libbux Jun 1 '13 at 6:01
  • 4
    Basically, yes. "Speed" in networking parlance is not so much how fast it is, but how many bits per second you can push over a link. Latency/RTT is usually a matter of the total distance combined with the number of devices in the path. Each hop/device adds a small amount of latency. The "fastness" is basically down to physics and the speed of light. – Justin Seabrook-Rocha Jun 1 '13 at 6:03
  • The key point there is that the bits all travel the same speed - that's really what I was wondering. – Libbux Jun 1 '13 at 6:07
  • 2
    I'll also mention that all this goes out the window when you enter the land of financial trading. In that world, RTT is king. Even microseconds of RTT matter, which causes them to do all sorts of silly things like move the workstation 100ft closer to the server and shorten up the fiber drop. – Justin Seabrook-Rocha Jun 1 '13 at 6:12
  • 7
    While signal travels at the same rate regardless if it's 100M, 1GE or 10GE, serialization delay (time it takes to inject frame to the wire) goes down as rate goes up. That is, if you send frame on 100M serialization delay for 1538B is 0.123ms, on 1GE that frame is already 21km down the fibre. – ytti Jun 1 '13 at 6:46
6

The speed of the flow of data makes no difference in the physics of the medium. By this I mean that it takes the same time for an electric signal to flow from one side of a 100 meter copper run to the other, no matter if that signal is part of a 10Mbps or a 1Gbps link.

If you change from copper to fiber, then you may notice a small improvement, but it really should be only a marginal difference.

Now, there are other factors that may come into play, for instance the equipment that can do 10Gbps is generally more capable of processing the frames/packets than equipment that is designed to do 10Mbps, so the latency added by the equipment may be reduced as well. But this is entirely dependent on the capabilities of the equipment and not on the speed of the link.

| improve this answer | |
5

In this case, moving from 1G end-to-end, to a 10G core shouldn't significantly change anything. Only a marginal increase in throughput would come from the faster signalling (decreased bit-time) on the 10G+ link. But in the absence of any congestion (read: other hosts), they should've been able to saturate the link to begin with.

The time it takes hosts A & B to signal (in and out) a packet doesn't change. The time it takes the packet to hop from switch to switch is, in theory, proportionally faster. However, at these speeds, the difference isn't noticeable to a human. (~10μs for 1500 mtu packet)

| improve this answer | |
  • Would, say, a 100G link between switches make a noticeable difference? – Libbux Jun 1 '13 at 23:36
  • They can already more than saturate their their local link, so no. With sensitive test equipment, you could measure the change in per-packet latency, but it won't make the one gig links go any faster. :-) That said, the real benefit is congestion avoidance when there are more than just host A and B using the link. – Ricky Beam Jun 1 '13 at 23:40
0

Since throughput is = to windows size / RTT anything that shortens RTT would increase throughput, it is a different question as to if it is worth it. The larger the window size the more impact decreasing RTT has.

| improve this answer | |
  • 1
    with a 128k window, 1G end-to-end would be ~3.5G/s. with a 10G link in the middle, that number jumps to ~5G/s. You'll note both numbers are well above the host link speed, so it would not be noticeable at all. – Ricky Beam Jun 1 '13 at 23:36
0

It depends.

In an otherwise idle network it depends on whether the switching devices are "store and forward" or "cut through". If the switching devices are store and forward then faster links will mean lower latency. However if they support cut-through switching then extra latency will be introduced as it is not possible to do cut through switching from a slower incoming link to a faster outgoing link. However unless you are playing in the high frequency trading world or similar this is likely to be negligable either way.

In a practical network having more capacity in the core decreases the chance that congestion from other users will be encountered. Congestion drives throughput down and latency up. In general it is good if your core links are faster than your end user links so no one end user can saturate them (so if you are running gigabit to the desktop you should probablly be running a 10 gigabit core).

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.