8 ms·

Propagation Delay and Its Relationship to Maximum Cable Length

78 points by andrelaszlo 14y ago

andrewvc 14y ago
Cool article. Though it only talks about ancient Ethernet variants. Issues like packet collision don't happen on most networks because we use switches instead of vampire taps and hubs. Can someone more knowledgeable explain the parts that are still correct?
- X-Istence 14y ago
  All of them. Just because Station A and Station B are switched means that we could now say that the Switch is Station C. So lets say Station A starts transmitting at the same time as Station C (our switch is sending a packet from Station B to Station A) we can still get a collision. We no longer get a collision between the nodes on the switch because of the fact that the switch will buffer packets as necessary before sending them on to their intended target, which reduces the amount of collisions that can/do occur...
  - mikeash 14y ago
    I thought this stuff was full-duplex now, with each direction using different wires. So there's only ever one device that can transmit on a given set of wires, thus no chance for collisions. Is that not the case?
    - smithian 14y ago
      You are correct. There is very little half-duplex around these days except in some carrier ethernet situations.
    - mprovost 14y ago
      Yes that is true for gigabit. Gigabit ethernet is ethernet in name only. It doesn't resemble the original standard in most meaningful ways. But the name is a brand at this point so whatever the new standard is, it's always called ethernet even if it's completely different electrically. Or even optically in the case of fibre, which is something that was never originally anticipated that it would run over.
      - yuhong 14y ago
        Backward compatiblity too.
thedufer 14y ago
Very misleading article title. The entire thing is about collisions, which only limit the length between ethernet cards, which is 5 times the length in the title. I suspect the 500m is related to the bandwidth and signal degradation rather than anything in the ethernet protocol.
tlb 14y ago
None of this applies to modern gigabit ethernet installations, which are full-duplex and switched so that collisions don't occur. Maximum cable length is limited by signal degradation, so the limit is soft. With good quality cables you can go a little beyond the rated maximum.
- ajross 14y ago
  Exactly. The switches buffer packets. The article is talking about a CSMA limitation of the old unbuffered hub architecture (and the "real electrical bus" coax networks before that) which have been irrelevant for over a decade now. And my experience is that you can actually get well beyond the official lengths for most signal types even with comparatively cheap cables as long as you're willing to test and throw out the odd cable or connector. Standards are written with conservatism in mind.
  - jcr 14y ago
    > Exactly. The switches buffer packets. In most cases, you're right. The vast majority of switches use a sore-and-forward design (buffering), but when you get into latency-critical applications, another type of switch design is used, typically called a cross-bar switch. http://www.researchgate.net/publication/220055485_A_low-latency_modular_switch_for_CMP_systems http://www.researchgate.net/publication/220055485_A_low-late... http://www.google.com/patents/US5339396 http://www.google.com/patents/US5339396 I think opinions vary on whether or not crossbar designs "buffer" but it mostly depends on how you define buffering. None the less, the low-latency crossbar designs try to eliminate the latency caused by buffering in the older but more common store-and-forward designs.
    - ajross 14y ago
      This is a valid point and I should have been more precise. But AFAICT all those interconnects are single duplex point-to-point links, and thus not subject to the CSMA issues of asymmetric collision detection detailed in the linked article.
- sliverstorm 14y ago
jacquesm 14y ago
Cat6 and cat5 are recommended to go no further than 90 meters: http://en.wikipedia.org/wiki/Category_6_cable http://en.wikipedia.org/wiki/Category_6_cable http://en.wikipedia.org/wiki/Category_5_cable http://en.wikipedia.org/wiki/Category_5_cable Links are full duplex nowadays, the bus architecture is long gone and hubs are rare (usually just old), switches are dirt cheap so no reason for hubs. The limiting factor is simply the ability of the electronics to recover the signal reliably. You can go way beyond the specs if you are willing to accept ever increasing packet losses due to degradation of the signal, one way to improve throughput over such links is by shortening your MTU. Another simple solution to make longer runs is to use a fibre-optic repeater (for instance this ancient 10 mbit device: http://www.blackbox.com/resource/files/productdetails/10542.PDF http://www.blackbox.com/resource/files/productdetails/10542.... but there are tons of others, more modern gear: http://www.signamax.com/mediaconverters/13-10100basettx-to-100basefx-single-fiber-converters http://www.signamax.com/mediaconverters/13-10100basettx-to-1... as an additional advantage, fibre optics don't suffer from electrical interference and act as an optical barrier for electrical signals which is very handy in industrial environments). Or to forego cabling entirely and to use wireless with directional antennae.
- zalew 14y ago
  > Or to forego cabling entirely and to use wireless with directional antennae. but then comes the weather factor.
  - jacquesm 14y ago
    Over links up to a few kilometers this is usually not a problem if you can get a line-of-sight setup.
    - zalew 14y ago
      A few years ago I've had one for my office, I don't remember the exact location but it was in the 2-5km range, terrible reliability during storms and snowfall.
      - 14y ago
jpdoctor 14y ago
> Since we know that Ethernet operates at 10Mbps or 10,000,000 bits per second I'm trying to figure out if this essay was written in the late 80s or the early 90s.
powertower 14y ago
> Electrical signals in a copper wire travel at approximately 2/3 the speed of light. An interesting fact - in those copper wires (or any type of wires) the electron velocity itself is less than 1 centimeter / second (and often in the mm/s range). Whether it's DC or AC, it does not matter much. Think of the wire as a tube with a bunch of ping-pong balls inside - filling it up tightly from start to end. If I stick my finger into one end of the tube, a ball will come out at the other end almost instantaneously (regardless of how long the tube is), but the balls themselves only moved a small distance - and only at the speed/velocity I pushed my finger at. But the "signal" "traveled" at a much higher velocity. So in DC - electrons only travel a few centimeters / second (not anywhere near light speed), and in AC - you're pushing/pulling the same electron back and forth for eternity at about the same speed as it travels in DC. Now it's of course a bit more complicated (the electrons rattle around at much greater speeds, and the mentioned velocity is a net average), and there are some edge cases, but for the most part the above is accurate. I'm just not 100% sure of the speed of electrons in a superconducting loop (Cooper-pair electrons) - if they are truly relativistic or not. And if this does not blow your mind, you should not be reading about electricity. PHDs in the electrical field don't even know this stuff. http://en.wikipedia.org/wiki/Drift_velocity http://en.wikipedia.org/wiki/Drift_velocity TL;DR; If you raced a snail and an electron in a wire together, the snail would win - https://www.youtube.com/watch?v=jbi7gJTPSXk https://www.youtube.com/watch?v=jbi7gJTPSXk
- greenmountin 14y ago
  One of the coolest things about superconductors is "kinetic inductance"[1]. There is no energy in a normal current. As you say, the electrons don't travel all that fast most of the time. They feel a field, accelerate until they bump into something and start all over again. That's resistance; the energy stored by an inductor or capacitor (a current or voltage) is in the magnetic or electric field nearby. This isn't true for Cooper-pairs in superconductors. They go fast enough that there is true energy stored in their velocity; there's nothing to stop them, they're a superfluid. This is used for certain types of detectors (MKIDs), and means that there is a bit of extra L in superconducting transmission lines, above the geometric inductance. [1] http://en.wikipedia.org/wiki/Kinetic_inductance http://en.wikipedia.org/wiki/Kinetic_inductance
jleader 14y ago
When was this article written? The last time I worked in an office with 10Base2 cabling (aka "thin ethernet", using RG-58 coax about 5mm thick with BNC connectors) was 1995. The last time I saw 10Base5 (aka "thick ethernet", using an extra-stiff variant of RG-8 coax about 10mm thick, in continuous runs with "vampire taps") was sometime in the late 1980s. And if the maximum cable length were simply proportional to signal propagation speed / bit rate, then why does 10Base2 have a shorter maximum length than 10Base5? They're both carrying 1e6 bits per second. Do signals really travel 3 times faster through fat coax than through thin?
smoyer 14y ago
Propagation delay is dependent upon a variety of factors and while the numbers the article quotes may be correct for thick-net and thin-net, the cable we use for CATV trunk-line is significantly faster at about 0.85c (http://www.china-satellite.com/sdp/1385167/4/pd-5853918/8382534-2286595/RG540_QR540_QR540JCA_QR540JCAM109_Trunk_Line_Cable.html http://www.china-satellite.com/sdp/1385167/4/pd-5853918/8382...). Interestingly, the propagation delay through single-mode fiber tends to be significantly lower than through copper (well ... again it depends on the cable), but the 0.66c speed is a pretty good number for fiber.
ghshephard 14y ago
"You may know that the minimum frame size in an Ethernet network is 64 bytes or 512 bits, including the 32 bit CRC. You may also know that the maximum length of an Ethernet cable segment is 500 meters for 10BASE5 thick cabling and 185 meters for 10BASE2 thin cabling. It is, however, a much less well known fact that these two specifications are directly related." Odd, given that the VERY first thing anybody is taught about CSMA/CD ethernet is how the maximum length of cable and minimum packet size are precisely related. Indeed, it's almost the only thing people are taught about the relationship of those two.
weichi 14y ago
Interesting article. A nit that can be picked is that it is not generally true that electrical signals in copper travel slower than c. The reason the signal travels slower than c in an ethernet cable is that the copper is surrounded by dielectric material, and it is the presence of the dielectric that slows down the signal. If the 10BASE5 coax cable was filled with air instead of foam, the signal would travel at c.

78 points by andrelaszlo 14y ago

Cool article. Though it only talks about ancient Ethernet variants. Issues like packet collision don't happen on most networks because we use switches instead of vampire taps and hubs. Can someone more knowledgeable explain the parts that are still correct?

X-Istence 14y ago
All of them. Just because Station A and Station B are switched means that we could now say that the Switch is Station C. So lets say Station A starts transmitting at the same time as Station C (our switch is sending a packet from Station B to Station A) we can still get a collision. We no longer get a collision between the nodes on the switch because of the fact that the switch will buffer packets as necessary before sending them on to their intended target, which reduces the amount of collisions that can/do occur...
- mikeash 14y ago
  I thought this stuff was full-duplex now, with each direction using different wires. So there's only ever one device that can transmit on a given set of wires, thus no chance for collisions. Is that not the case?
  - smithian 14y ago
    You are correct. There is very little half-duplex around these days except in some carrier ethernet situations.
  - mprovost 14y ago
    Yes that is true for gigabit. Gigabit ethernet is ethernet in name only. It doesn't resemble the original standard in most meaningful ways. But the name is a brand at this point so whatever the new standard is, it's always called ethernet even if it's completely different electrically. Or even optically in the case of fibre, which is something that was never originally anticipated that it would run over.
    - yuhong 14y ago
      Backward compatiblity too.

thedufer 14y ago

Very misleading article title. The entire thing is about collisions, which only limit the length between ethernet cards, which is 5 times the length in the title. I suspect the 500m is related to the bandwidth and signal degradation rather than anything in the ethernet protocol.

tlb 14y ago

None of this applies to modern gigabit ethernet installations, which are full-duplex and switched so that collisions don't occur. Maximum cable length is limited by signal degradation, so the limit is soft. With good quality cables you can go a little beyond the rated maximum.

ajross 14y ago
Exactly. The switches buffer packets. The article is talking about a CSMA limitation of the old unbuffered hub architecture (and the "real electrical bus" coax networks before that) which have been irrelevant for over a decade now. And my experience is that you can actually get well beyond the official lengths for most signal types even with comparatively cheap cables as long as you're willing to test and throw out the odd cable or connector. Standards are written with conservatism in mind.
- jcr 14y ago
  > Exactly. The switches buffer packets. In most cases, you're right. The vast majority of switches use a sore-and-forward design (buffering), but when you get into latency-critical applications, another type of switch design is used, typically called a cross-bar switch. http://www.researchgate.net/publication/220055485_A_low-latency_modular_switch_for_CMP_systems http://www.researchgate.net/publication/220055485_A_low-late... http://www.google.com/patents/US5339396 http://www.google.com/patents/US5339396 I think opinions vary on whether or not crossbar designs "buffer" but it mostly depends on how you define buffering. None the less, the low-latency crossbar designs try to eliminate the latency caused by buffering in the older but more common store-and-forward designs.
  - ajross 14y ago
    This is a valid point and I should have been more precise. But AFAICT all those interconnects are single duplex point-to-point links, and thus not subject to the CSMA issues of asymmetric collision detection detailed in the linked article.
sliverstorm 14y ago

jacquesm 14y ago

Cat6 and cat5 are recommended to go no further than 90 meters: http://en.wikipedia.org/wiki/Category_6_cable http://en.wikipedia.org/wiki/Category_6_cable http://en.wikipedia.org/wiki/Category_5_cable http://en.wikipedia.org/wiki/Category_5_cable Links are full duplex nowadays, the bus architecture is long gone and hubs are rare (usually just old), switches are dirt cheap so no reason for hubs. The limiting factor is simply the ability of the electronics to recover the signal reliably. You can go way beyond the specs if you are willing to accept ever increasing packet losses due to degradation of the signal, one way to improve throughput over such links is by shortening your MTU. Another simple solution to make longer runs is to use a fibre-optic repeater (for instance this ancient 10 mbit device: http://www.blackbox.com/resource/files/productdetails/10542.PDF http://www.blackbox.com/resource/files/productdetails/10542.... but there are tons of others, more modern gear: http://www.signamax.com/mediaconverters/13-10100basettx-to-100basefx-single-fiber-converters http://www.signamax.com/mediaconverters/13-10100basettx-to-1... as an additional advantage, fibre optics don't suffer from electrical interference and act as an optical barrier for electrical signals which is very handy in industrial environments). Or to forego cabling entirely and to use wireless with directional antennae.

zalew 14y ago
> Or to forego cabling entirely and to use wireless with directional antennae. but then comes the weather factor.
- jacquesm 14y ago
  Over links up to a few kilometers this is usually not a problem if you can get a line-of-sight setup.
  - zalew 14y ago
    A few years ago I've had one for my office, I don't remember the exact location but it was in the 2-5km range, terrible reliability during storms and snowfall.
    - 14y ago

jpdoctor 14y ago

> Since we know that Ethernet operates at 10Mbps or 10,000,000 bits per second I'm trying to figure out if this essay was written in the late 80s or the early 90s.

powertower 14y ago

> Electrical signals in a copper wire travel at approximately 2/3 the speed of light. An interesting fact - in those copper wires (or any type of wires) the electron velocity itself is less than 1 centimeter / second (and often in the mm/s range). Whether it's DC or AC, it does not matter much. Think of the wire as a tube with a bunch of ping-pong balls inside - filling it up tightly from start to end. If I stick my finger into one end of the tube, a ball will come out at the other end almost instantaneously (regardless of how long the tube is), but the balls themselves only moved a small distance - and only at the speed/velocity I pushed my finger at. But the "signal" "traveled" at a much higher velocity. So in DC - electrons only travel a few centimeters / second (not anywhere near light speed), and in AC - you're pushing/pulling the same electron back and forth for eternity at about the same speed as it travels in DC. Now it's of course a bit more complicated (the electrons rattle around at much greater speeds, and the mentioned velocity is a net average), and there are some edge cases, but for the most part the above is accurate. I'm just not 100% sure of the speed of electrons in a superconducting loop (Cooper-pair electrons) - if they are truly relativistic or not. And if this does not blow your mind, you should not be reading about electricity. PHDs in the electrical field don't even know this stuff. http://en.wikipedia.org/wiki/Drift_velocity http://en.wikipedia.org/wiki/Drift_velocity TL;DR; If you raced a snail and an electron in a wire together, the snail would win - https://www.youtube.com/watch?v=jbi7gJTPSXk https://www.youtube.com/watch?v=jbi7gJTPSXk

greenmountin 14y ago
One of the coolest things about superconductors is "kinetic inductance"[1]. There is no energy in a normal current. As you say, the electrons don't travel all that fast most of the time. They feel a field, accelerate until they bump into something and start all over again. That's resistance; the energy stored by an inductor or capacitor (a current or voltage) is in the magnetic or electric field nearby. This isn't true for Cooper-pairs in superconductors. They go fast enough that there is true energy stored in their velocity; there's nothing to stop them, they're a superfluid. This is used for certain types of detectors (MKIDs), and means that there is a bit of extra L in superconducting transmission lines, above the geometric inductance. [1] http://en.wikipedia.org/wiki/Kinetic_inductance http://en.wikipedia.org/wiki/Kinetic_inductance

jleader 14y ago

When was this article written? The last time I worked in an office with 10Base2 cabling (aka "thin ethernet", using RG-58 coax about 5mm thick with BNC connectors) was 1995. The last time I saw 10Base5 (aka "thick ethernet", using an extra-stiff variant of RG-8 coax about 10mm thick, in continuous runs with "vampire taps") was sometime in the late 1980s. And if the maximum cable length were simply proportional to signal propagation speed / bit rate, then why does 10Base2 have a shorter maximum length than 10Base5? They're both carrying 1e6 bits per second. Do signals really travel 3 times faster through fat coax than through thin?

smoyer 14y ago

Propagation delay is dependent upon a variety of factors and while the numbers the article quotes may be correct for thick-net and thin-net, the cable we use for CATV trunk-line is significantly faster at about 0.85c (http://www.china-satellite.com/sdp/1385167/4/pd-5853918/8382534-2286595/RG540_QR540_QR540JCA_QR540JCAM109_Trunk_Line_Cable.html http://www.china-satellite.com/sdp/1385167/4/pd-5853918/8382...). Interestingly, the propagation delay through single-mode fiber tends to be significantly lower than through copper (well ... again it depends on the cable), but the 0.66c speed is a pretty good number for fiber.

ghshephard 14y ago

"You may know that the minimum frame size in an Ethernet network is 64 bytes or 512 bits, including the 32 bit CRC. You may also know that the maximum length of an Ethernet cable segment is 500 meters for 10BASE5 thick cabling and 185 meters for 10BASE2 thin cabling. It is, however, a much less well known fact that these two specifications are directly related." Odd, given that the VERY first thing anybody is taught about CSMA/CD ethernet is how the maximum length of cable and minimum packet size are precisely related. Indeed, it's almost the only thing people are taught about the relationship of those two.

weichi 14y ago

Interesting article. A nit that can be picked is that it is not generally true that electrical signals in copper travel slower than c. The reason the signal travels slower than c in an ethernet cable is that the copper is surrounded by dielectric material, and it is the presence of the dielectric that slows down the signal. If the 10BASE5 coax cable was filled with air instead of foam, the signal would travel at c.