4.2 What's inside a router? Part 2.

[Music] well now that we've got a really good appreciation for the pieces of the router the input ports the output ports and the switching fabric we're going to dive down deep into the topics of packet scheduling and buffer management and we've seen how important these are with the Internet's best effort service model packet loss and packet delay have had a profound effect on what has to happen the transport layers and the application layers so now we're going to dive down deep into the topics of packet scheduling and buffer management and we'll take a look at

how those can be used to control and also mitigate packet loss and packet delay well let's get started by first taking a look at input port queuing remember that we saw earlier that input port queuing can happen when the routers switching fabric is less than the combined rates of its incoming links in this case more packets can arrive in a unit time then the switching fabric can possibly move from the input side of the fabric to the output side and so queuing can occur on the input side a queuing phenomena that's unique to the input

side of the switching fabric is what's known as head of the line blocking or hol end of the line blocking occurs when packets at different input ports want to go to the same output port if only one pack at a time can be transferred from an input to a given output port and if there are K packets that want to go to the same output port then K minus 1 of those packets are going to have to wait on the input side of the switch and if packets have to wait they can also block the

packets behind them in their input queue from moving forward and this is what's known as head of the line blocking take a look at this example here where the colors of the packets correspond to the output port to which they're destined in this example say at time T we see the top and bottom input ports have a red packet destined to the upper output port in one clock deck the upper red packet is transferred from the input to the output side of the switch the blue packet in the middle it's also able to move from

its input port to its output port the red packet at the bottom however is locked from moving to the read output port since the packet at the top was already transferred there and so the green packet behind this lower red packet experiences head of the line blocking let's now focus on the output port side since this is really where the action is and let's start by seeing why buffering is needed in the first place at this output port we have a link with transmission rate of our bits per second and let's say we have a

non-blocking switch capable of delivering end times our bits per second to the output port now we can see why buffering occurs bits can be arriving at a rate of end times are but bits can only be drained or transmitted out at a rate of are whenever the arrival rate exceeds the departure rate the buffer will fill and since buffers are finite there may not be enough buffer space to store all of the packets that need buffering and packets will have to be dropped there simply isn't enough memory to hold all of the packets that need

buffering so right here at the output port is where congestive packet loss occurs and you know it may be tempting to say well it's really the switch fabrics fault it's delivering packets too fast at a rate of end times are when the link can only drain at rate R so we should slow the fabric down but that's not really the cause of the problem is it think back to our earlier discussion of principles of congestion control the real cause of the problem congestion is that there are too many transport layer senders at the edge of

the network sending too many packets too fast packets whose paths cross at this congested output port in the face of buffering and congestion there are a couple of decisions that'll need to be made at the output port first if there's not enough room for all arriving packets a packet drop policy will be needed to determine which packets are going to be dropped and if there are multiple packets in the buffer awaiting transmission a packet scheduling discipline will be needed to choose which packets among the many cute packets will be transmitted next we'll look at both

drop and scheduling policies in just a second now there are a lot of really important idea that we just talked about here on this slide so make sure you've got it all it's really important well you might be thinking given buffering is such a big concern how much buffering should a router have and more to the point you might also be thinking well hey shouldn't it be a good idea to invest in a lot of buffering with a lot of buffering there'd be less packet loss since there'll be more buffers to observe the burstiness and

packet flows having more buffering must be a good thing right well it might surprise you that even after 50 years the question of how much buffering is the right amount of buffering is still pretty much an unresolved issue there are some recommendations you could take a look at RFC 3439 and find a rule of thumb that says the amount of buffering should equal a typical RTT say a couple hundred milliseconds times the output link capacity there's also been more a recent theoretical research that assumes that transport layer senders are generally independent of each other and

recommends dividing the RFC 34:39 amount of buffering by the square root of n where n is the number of flows crossing the link and that's a lot less buffering and while you might have the sense that large buffers will generally result in less loss and you could be right large buffers do have a downside large buffers mean larger delays and larger RT T's large delays are bad for gamers and for interactive teleconferencing users where tens of milliseconds really do matter but maybe even more importantly large RTT delays mean that TCP senders will be slower to

detect and therefore react to congestion you might recall that when we studied delay based congestion control we encountered the saying that we want to keep the bottleneck link just full enough to keep the link busy but no fuller but that was from the perspective of the sending hosts when we think about the amount of buffering we really do want buffers to absorb short term statistical fluctuation in packet arrival rates and to keep the link busy but not to have so much buffering that congestion control action times become slow and sluggish well buffering sort of like

salt and cooking you've got to find the right amount just the right amount makes things better but too much can really ruin things well who would have thought that buffering and queuing at an output port could actually be so complicated and so subtle and you may ask yourself why is that and it's really because it's here at an output port buffer that the global behavior of potentially thousands of active senders worldwide are converging as their flows all pass through the single output link we see the convergence of global internet scale behavior at this one location

deeply inside the internet that's pretty cool well we're next going to move on to the mechanisms for buffer management and packet scheduling and will find it useful to think of an output port abstractly as a queue as shown here of course we're all familiar with queuing systems in our daily lives banks at buses at bus stops at food stores at metros and many other places think about how many cubes you've been in today packets arrive to the queue they wait for service and they're eventually selected for service receive service and then leave the queue the

link layer functions as the server when a packets bits are being transmitted over the outgoing link that packets in service if an arriving packet emerges from the switching fabric to find there's no free buffer space either that packets going to be dropped and that's known as tail drop or an already queued packet could be removed from the buffer and dropped to make room for this newly arriving packet say according to some kind of priority mechanism for example there might be a policy to drop low priority packets say carrying end-user TCP traffic rather than high priority

packets say carrying Network management in control packets for the ISP as we also learned when we studied Network assisted congestion control a buffered packet may also be marked with a congestion indication to signal that the link is becoming congested we learned earlier that for the internet's explicit congestion notification ecn mechanism to bits in the IP header type of service field are used for ecn marking will see these bits in the IP datagram header in the next section when we study ipv4 but it's here at the output ports and where congestion occurs that these bits are

going to be set we can now look at several packet scheduling disciplines that are used in practice under first come first serve scheduling packets are transmitted simply in the order in which they arrive to the output port we're all familiar with first-come first-serve queuing I'm guessing it's what people do the most we all join the back of the queue and customers are served from the front of the queue same thing with packets we're all familiar with priority queues you get that with Airlines platinum diamond gold silver or bronze class of service they all sound pretty

fancy and great but you definitely rather be platinum than bronze and I think we all probably have an intuitive understanding of how priority disciplines work having been in the lower priority class of service and seeing later arriving people in higher priority classes getting service before us under priority scheduling packets arriving at the output link are classified into priority classes on arrival at the queue as shown here where we have a high priority red class of traffic and a lower priority green class of traffic the priority queueing discipline will transmit a packet from the highest priority

class that has a non-empty queue that is has packets waiting for transmission the choice among packets in that same priority class is typically done in a first-come first-serve manner this animation shows packets arriving from the top entering service according to a priority discipline and then leaving the queue in this example packet one arrives finds the queue empty and enters service while packet one is in service packet two and three arrived when packet one finishes both packets two and three are buffered but packet three is chosen for transmission even though it arrived after packet two since

it has higher priority than back at two and so on well we've now learned about the priority queuing mechanism but you may still be wondering what is it that defines a priority class in the first place how is it determined what goes into priority class 1 or priority class 2 for example and this is actually determined by the network operator by the ISP the ISP may decide that well network management traffic's really important it's the highest priority class voice-over-ip gets higher priority than email and so on and you remember that by looking at the transport

layer port numbers one can determine what type of traffic is being carried by that data graph now traffic type is one way in which priorities could be determined another way might be on the basis of source or even destination addresses so here's another possible way to classify traffic into priority classes imagine that a company is willing to pay an internet service provider to give its packets better service packets are less likely to be dropped its packets suffer less delay as a result if it's a search company its search results get to users faster if it's

voice over IP the voice is more responsive if it's streaming video there's less spinning meals well we've got the scheduling mechanisms to do that whether or not a company is able to pay for that kind of better service actually gets us into the topic of what's known as network neutrality we're going to cover that in just a second but first let's wrap up by looking at a few more scheduling disciplines so we've seen first-come first-serve and priority scheduling under round-robin scheduling packets are assigned to classes as with priority scheduling however rather than there being a

strict service priority among classes a round-robin scheduler alternates service among classes for example a round-robin scheduler would choose a packet from class 1 followed by a packet from class 2 followed by a packet from class 3 and so on and if there's no packet of a given class the scheduler moves on to the next class in the schedule a generalized form of round-robin queuing that's been widely implemented in routers and practice is known as weighted fair queuing wfq light priority and round-robin scheduling arriving packets are classified and queued in the appropriate per class waiting area

we see here red green and blue packets waiting for their service a weighted fair queuing scheduler schedules classes in a round-robin like manner first serving class 1 then serving class 2 then serving class 3 and then assuming there are just three classes repeating that service pattern weighted fair queuing differs from round-robin in that each class can receive a different amount of service in any interval of time and here's how it works ideally each class of packets I has a weight W sub I let's assume that all of the weights the sum of W J say

sum up to 1 under weighted fair queuing during any interval of time during which there are class I packets to send class I will then be guaranteed to receive a fraction of service W sub I that means if the link has a rate R then each class of service will get a guaranteed minimum amount of bandwidth W sub I times R and so we can see how weighted fair queuing allows some type of a bandwidth guaranteed to be made on a per class basis network neutrality is an area where our technical interests in computer networking

intersect with social political and economic considerations well we've already seen the mechanisms for sharing resources but what about the social political and economic considerations about how those resources are shared let's take a look network neutrality is about the laws and policies that govern how an Internet service provider can use the traffic control mechanisms that we've just learned about packet scheduling in particular to control network traffic so what are some of the questions that network neutrality addresses well it's about a number of different things it's about free speech for example cannon Espie refused to carry some

types of traffic for example certain types of news or political opinion it's also about encouraging innovation and about encouraging competition for example must an isp give all companies big and small equal treatment of their traffic as you probably know different countries have been adopting different stances with respect to these questions we're going to take a look at what's happening in the united states since this is well documented and has all been playing out pretty publicly our textbooks got some interesting references about net neutrality worldwide that you might want to take a look at in 2015

the United States Federal Communications Commission order on protecting and promoting an open Internet defined three clear bright line rules associated with network neutrality let's take a look at what they are the first bright line rule is no blocking and here's what it says it says that an isp quote unquote shall not block lawful content applications services or non-harmful devices subject to reasonable network management in quote now there's a lot packed in there the word lawful and applications services and devices well this rules about blocking traffic and there actually was a case involving an ISP that

had a practice of blocking its customers from using vonage which is a voice over IP service that competed with this is p's own telephone service that would not be allowed and notice the phrase subject to reasonable network management will cover network management when we cover the network layer control plane but note here that it is allowable to block traffic for network management purposes say if the networks been degraded and then only trickles of traffic can get through then it's best that network management traffic be able to get through of course the question of what reasonable

network management is is left open to interpretation the second bright-line rule is no throttling that is an isp quote unquote shall not impair or degrade lawful Internet effect on the basis of Internet content application or service or use of a non harmful device again subject to reasonable network management note here the words impaired or degrade there was a case not too long ago where an ISP was judged to be interfering with BitTorrent peer-to-peer traffic and it was doing this by internally creating and sending TCP reset packets to BitTorrent client and servers executing at the end

host this caused the host to close their BitTorrent connections basically the ISP was closing the TCP connection underneath the BitTorrent application whose traffic it really didn't want and then finally there's the bright-line rule of no paid prioritization this means for example that video streaming traffic from all video streaming service providers must be treated the same for example that one streaming video service provider cannot pay to get its packets better service on their way to the customer and of course we've seen how this can be done from a technical point of view via priority queuing well

you might ask why no paid prioritization since paying for a priority service seems well almost like a fact of life at least here in the United States well one of the arguments is that if say a video streaming service provider can pay for better service and if that successful incumbent the existing service provider can pay a lot then this introduces a high barrier to entry for new competitors and of course some countries want to encourage market competition but there are counter arguments to this claiming that creating additional revenues for an ISP by having paid prioritization

is a good thing from a competitive point of view since that makes the ISP marketplace more lucrative and so more companies will want to invest in the ISP market or ISPs will want to invest more in their infrastructure and then finally there are laws going back almost a hundred years that have regulated the telephone network as a telecommunications service provider but there's much less regulation of information service providers so what is an ISP is it a telecommunication service provider or an information service provider well it turns out from a regulatory standpoint the answer to this

question matters a lot in the US but the answer to this question has really not yet been finally determined well these bright-line rules are from a 2015 u.s. FCC order more recent FCC orders have actually rewritten much of that 2015 order so until laws are written and until there's further litigation and case law adopted I think it's fair to simply say that the situation continues to evolve and I think that's true in many countries it's interesting that we've had the ability to do prioritization since the internet first became public nearly 30 years ago so although

the Internet has a strong technical foundation as we've seen three decades later our social political and economic policies governing this creation that have not yet been fully defined or agreed upon well that wraps up our discussion of packet scheduling and packet buffering and also what I hope was an interesting sidebar into the topic of network neutrality and more generally it wraps up our discussion of what's inside a router remember we also looked at input ports output ports and switching fabrics coming up next we're going to take a look at the Internet Protocol ipv4 so stay

tuned for that you [Music]