5 ms·

slightly tangent: is it true that ethernet doesn't have a solid math theory behind? if so, how does it scale?

by mshockwave 5y ago

markjenkinswpg 5y ago
Switching.
gowld 5y ago
Why would math theory be required for scaling?
- fragmede 5y ago
  Implementation details at layer 1 and 2 of the OSI model. If machine aa needs to talk to machine zz (and there are all the machine in between them), then aa needs a timeslice to change the electrical levels that are on the actual ethernet cable, as does zz in order to reply, but if ab and ac and ad and all the way to zy are also busy communicating, then when does zz get that timeslice? Switching (vs hubs) improves the situation drastically, but as you scale up (like, thousands of thousands aka "webscale", there's clearly going to be some sort of limit at some point (until you apply other technology). What is that point using Ethernet though, and how do you calculate that?
  - thehappypm 5y ago
    The OSI model never really panned out in reality
    - birdyrooster 5y ago
      Except it did, and that is the point of Layer 3. Also the esoteric nature of OSI doesn't mean it isn't useful.
      - thehappypm 5y ago
        What I mean is that the OSI layered model never got implemented exactly to spec.
        birdyrooster 5y ago
        Yes good point
justin66 5y ago
Of course it's not true.
gjf 5y ago
Short answer is: it doesn't. When designing networks you will try to reduce the scope of layer 2 broadcast domains.
- NovemberWhiskey 5y ago
  Right - the old rule of thumb used to be an Ethernet collision domain should have a maximum of five segments, with four repeaters, and three populated mixing segments (i.e. buses like 10BASE5 or 10BASE2), the other two segments being link segments (i.e. point-to-point like 10BASE-T or 10BASE-FL). I think the idea was that anything more than this was likely to lead to long enough round-trip times between devices at the far ends of the Ethernet that the CSMA/CD algorithm couldn't be guaranteed to work.
adrianmonk 5y ago
I'm old enough to have been a sysadmin before ethernet switches were available. We used (literal) hubs, where every packet goes to every machine. And it really did not scale. If you got too many computers on one Ethernet network together, things would start to really bog down. We would address this by splitting an ethernet network into two. When it got too slow, we'd take our network of (say) 1.2.3.0 - 1.2.3.255 and split it into two networks, 1.2.3.0 - 1.2.3.127 and 1.2.3.128 - 1.2.3.255. Each of these ethernet networks would get its own separate port in the IP router. (Which could require adding cards to the router. And put extra processing load on it.) DHCP did not exist yet, so we had to manually go to each computer and change its broadcast address, subnet mask, and default route. This was disruptive, so we tried to keep all addresses in the lower portion of the address range so that they would continue to function well enough during such a transition. For example, a host at 1.2.3.10 with subnet mask 255.255.255.0, router address 1.2.3.1, and broadcast address 1.2.3.255 would still function when we changed its ethernet to be 1.2.3.0 - 1.2.3.127 instead of 1.2.3.0 - 1.2.3.255. Its broadcast address and netmask were wrong, but it wasn't completely offline. (By contrast, a host with address 1.2.3.200 would be out of range for the new network parameters and would have problems immediately.) Then Grand Junction came up with the idea of ethernet switches to isolate traffic, and we didn't get one because they were a bajillion dollars. I don't know the exact prices, but I would guesstimate something like $5000 to $10000 for a switch with around 10 ports. But eventually they came down in price. With an ethernet switch, everything was amazingly easy in comparison. It basically did the same thing I'm talking about (splitting an ethernet into pieces so the devices don't try to talk over each other), except instead of it being a weeks-long manual process, it happened automatically in real time. Once switches became standard, ethernet was pretty scalable. Broadcast traffic still isn't scalable if it's used too much, but otherwise it basically works fine.
- Melatonic 5y ago
  Literal hubs still generally beat the hell out of dealing with serial connections :-D
- adrianmonk 5y ago
  Replying to myself: I guess Grand Junction didn't come up with it. Probably Kalpana did? Grand Junction was just the first ethernet switch I heard of. (Also, ethernet bridges were a thing at the time, and they are basically 2-port switches. But arguably a switch with a bunch of ports was a different way of thinking about how to do an Ethernet network.)
birdyrooster 5y ago
It does not scale. That is why we have IP.

slightly tangent: is it true that ethernet doesn't have a solid math theory behind? if so, how does it scale?

markjenkinswpg 5y ago

Switching.

gowld 5y ago

Why would math theory be required for scaling?

fragmede 5y ago
Implementation details at layer 1 and 2 of the OSI model. If machine aa needs to talk to machine zz (and there are all the machine in between them), then aa needs a timeslice to change the electrical levels that are on the actual ethernet cable, as does zz in order to reply, but if ab and ac and ad and all the way to zy are also busy communicating, then when does zz get that timeslice? Switching (vs hubs) improves the situation drastically, but as you scale up (like, thousands of thousands aka "webscale", there's clearly going to be some sort of limit at some point (until you apply other technology). What is that point using Ethernet though, and how do you calculate that?
- thehappypm 5y ago
  The OSI model never really panned out in reality
  - birdyrooster 5y ago
    Except it did, and that is the point of Layer 3. Also the esoteric nature of OSI doesn't mean it isn't useful.
    - thehappypm 5y ago
      What I mean is that the OSI layered model never got implemented exactly to spec.
      - birdyrooster 5y ago
        Yes good point

justin66 5y ago

Of course it's not true.

gjf 5y ago

Short answer is: it doesn't. When designing networks you will try to reduce the scope of layer 2 broadcast domains.

NovemberWhiskey 5y ago
Right - the old rule of thumb used to be an Ethernet collision domain should have a maximum of five segments, with four repeaters, and three populated mixing segments (i.e. buses like 10BASE5 or 10BASE2), the other two segments being link segments (i.e. point-to-point like 10BASE-T or 10BASE-FL). I think the idea was that anything more than this was likely to lead to long enough round-trip times between devices at the far ends of the Ethernet that the CSMA/CD algorithm couldn't be guaranteed to work.

adrianmonk 5y ago

I'm old enough to have been a sysadmin before ethernet switches were available. We used (literal) hubs, where every packet goes to every machine. And it really did not scale. If you got too many computers on one Ethernet network together, things would start to really bog down. We would address this by splitting an ethernet network into two. When it got too slow, we'd take our network of (say) 1.2.3.0 - 1.2.3.255 and split it into two networks, 1.2.3.0 - 1.2.3.127 and 1.2.3.128 - 1.2.3.255. Each of these ethernet networks would get its own separate port in the IP router. (Which could require adding cards to the router. And put extra processing load on it.) DHCP did not exist yet, so we had to manually go to each computer and change its broadcast address, subnet mask, and default route. This was disruptive, so we tried to keep all addresses in the lower portion of the address range so that they would continue to function well enough during such a transition. For example, a host at 1.2.3.10 with subnet mask 255.255.255.0, router address 1.2.3.1, and broadcast address 1.2.3.255 would still function when we changed its ethernet to be 1.2.3.0 - 1.2.3.127 instead of 1.2.3.0 - 1.2.3.255. Its broadcast address and netmask were wrong, but it wasn't completely offline. (By contrast, a host with address 1.2.3.200 would be out of range for the new network parameters and would have problems immediately.) Then Grand Junction came up with the idea of ethernet switches to isolate traffic, and we didn't get one because they were a bajillion dollars. I don't know the exact prices, but I would guesstimate something like $5000 to $10000 for a switch with around 10 ports. But eventually they came down in price. With an ethernet switch, everything was amazingly easy in comparison. It basically did the same thing I'm talking about (splitting an ethernet into pieces so the devices don't try to talk over each other), except instead of it being a weeks-long manual process, it happened automatically in real time. Once switches became standard, ethernet was pretty scalable. Broadcast traffic still isn't scalable if it's used too much, but otherwise it basically works fine.

Melatonic 5y ago
Literal hubs still generally beat the hell out of dealing with serial connections :-D
adrianmonk 5y ago
Replying to myself: I guess Grand Junction didn't come up with it. Probably Kalpana did? Grand Junction was just the first ethernet switch I heard of. (Also, ethernet bridges were a thing at the time, and they are basically 2-port switches. But arguably a switch with a bunch of ports was a different way of thinking about how to do an Ethernet network.)

birdyrooster 5y ago

It does not scale. That is why we have IP.