ud710 ยป



Lamport's clock gave a fundamental ordering mechanism for events in a distributed system. This theoretical basis is all the more important in this day and age, when so many everyday services, email, social networks, e-commerce, and now even online education are becoming distributive. Incidentally, this is the same Lamport who gave us a way to order memory accesses in a shared memory multi-processors, through the sequential consistency memory model. In the next part of the lesson we turn our attention to more practical matters in distributed systems. Specifically, given that network communication is the key to performance for distributed services, the operating system has to strive hard to reduce the latency incurred in the system software for network services. Lamport's clock serves as the theoretical underpinning for achieving deterministic execution in a distributed system, despite the non determinism that exists due to the vagaries of a network. In this lesson, we will discuss techniques for making the operating system softwares tag efficient for network communication. Both by looking at application interface to the kernel as well as inside the kernel in the protocol stack itself, but first a quiz.


Lets say it takes me one minute to go from my office to the classroom, but the hallway is wide enough that five of us can walk side-by-side going to my office to the classroom. So the question to you is to illustrate the difference between latency and throughput. The question is, what is the latency incurred by me to get to the classroom? And the second part of the question is asking you, what is the throughput achieved if I walk with four other students side-by-side to get to the classroom?


This is just a fun quiz to get you thinking about latency and throughput. I am sure that most of you would have gotten that, the time to get to the classroom, after all, from my office is one minute, so that's the latency I'm going to observe, getting to my classroom. The interesting thing is the throughput. The throughput is five per minute, because the hallway is wide enough. For five of us to walk side by side, so the throughput that we can get is five per minute. The important thing I want you to see is that, the latency is not the inverse of throughput. On other words, if tomorrow I widen the hallway to make 10 people walk side by side to get to the classroom from my office, that's going to increase the throughput, but it does nothing to the latency, the latency is always going to be the time it takes for me to get from my office, to the classroom.

Latency vs Throughput

So it's important to understand these two concepts of latency and throughput. Latency is the elapsed time from an event. If it takes me one minute to walk from my office to the classroom, that's the latency that I'm going to experience for that event of walking from my office to the classroom. That's the elapsed time. Throughput is the number of events that can be executed per unit time. Bandwidth is a measure of throughput. So once again with this analogy of walking to the classroom from my office, if the hallway is wide enough to allow five, ten of us to walk in parallel side by side to the classroom, increases the throughput but it does nothing to the latency. The latency is going to be determined by how fast I can walk from my office to the classroom. So, the difference between latency and, and throughput is very important to understand. In other words, I can increase the bandwidth and that'll improve the throughput but it is not going to do anything to the latency itself. So in other words, higher bandwidth does not necessarily imply lower latency. You'll work hard to lower the latency. RPC is a basis for client server based distributed systems. And performance of RPC is crucial, specifically in the context of this lesson. Latency refers to the time it takes for an application generated message to reach it's destination. So for instance, if you're doing an RPC call from a client to the server, then the RPC call entails sending the argument from the client to the server. And there is work to be done here, work to be done in sending the message, work to be done here before the server can actually execute the server procedure. So it's the latency that we are concerned about. And what we will see is all the software components that comprise the latency for RPC based communication. And performance of RPC is very crucial in building client server systems. There are two components to the latency that is observed for message communication in a distributive system. The first component is the hardware overhead and the second component is the software overhead. The hardware overhead is really dependent on how the network is interfaced to the computer. So, typically in any computer, what you have is a network controller that interfaces the network to the CPU. And typically, the network controller operates by moving the bits of the message from the system memory of the node into it's private buffer, which is inside the network controller. And this part of it, moving the bits from the memory of the node into the internal buffer of the network controller is accomplished using what is called direct memory access. Meaning the network controller is smart enough to move the bits directly using the bus that connects the memory to the network controller without the intervention of the CPU. And this is what is called dir, direct memory access. And that's how the bits are moved from the memory of the system into the buffer of the network controller, and once it comes here, the network controller can then put the bits out on the wire, and this is where the bandwidth that you have connecting your node to the network comes into play. But there are also other types of network controllers where the CPU may actually be involved in the data movement, and in that case, the CPU does program I/O to move the bits from the memory into the buffer of the network controller, from which the network controller will then put it out on the network. But modern network controllers tend to be built using DMA technique, meaning that the network controller, once the CPU tells the network controller were in memory the messages to be sent on the wire, network controller does the rest in terms of moving the bits into it's internal buffer, and then from the buffer putting it out onto the network. The software overhead is what the operating system tax on to the hardware overhead of moving the bits out onto the network. So the latency, if you think about the latency as a whole for doing a network transmission, there is the software overhead incurred in the layers of the operating system to make the message available in the memory of the processor, ready for transmission. Once it is ready for transmission, the hardware overhead kicks in, and the hardware, the network controller in particular, moves the bits from the memory into it's buffer and then out on the wire. The focus of course being an operating system designers work, is to reduce the software overhead and take what the hardware gives you and think about how you can reduce the software overhead so that we can overall reduce the latency involved in transmission. Which is a sum of the hardware overhead and software overhead.

Components of RPC Latency

Let's now discuss the components of the RPC latency. By now we are all familiar with the semantics of RPC, namely, in RPC the client, is making a remote procedure call to a server, and it has to send the arguments of the call to the server so that the server can execute the server procedure and return the results back to the client. So if you look at the components of the RPC latency, it starts with the client, with a client call. So the client call subsumes a number of things. Number one, it is setting up the arguments for the call. The client has to set up the arguments for the procedure call. And then it makes a call into the kernel. And once the kernel is being informed that it wants to make this call, the kernel validates the call, and then marshals the arguments into a network packet, and sets up the controller to actually do the network transmission. That entire set of activities that the client program and the kernel is involved in, in getting ready a network packet to send out, is subsumed in this one line, which I say is the client call. The second part of the latency is the controller latency, and this is the part where the controller says, well, there is a message to be sent out. I know where it is in memory. I have to first DMA that message into my buffer and then put the message out on the wire. That's the controller latency and so this part of it is in hardware, and as operating system designer we're going to take what the hardware gives us. Controller latency is what you have, that given by the hardware. The third part of the latency is the time on the wire. Now this really depends, as one might imagine, on the distance between the client and the server. The limiting factor of course is speed of light. So, depending on the bandwidth that's available between the source and the destination, perhaps if you have to go through intermediate routers and so on. It is going to take a certain amount of time to go from the client to the server machine, and that we call as the time on the wire. So then, the message arrives over on the destination node, and it arrives in the form of an interrupt to the operating system. So, the interrupt has to be handled by the operating system, and part of handling the interrupt is moving the bits that come in on the wire into the controller buffer. And from the controller buffer into the memory of the node. So all of that activity is subsumed in this item number four, which I call the interrupt handling. So once the interrupt handling is complete, then we can set up the server procedure to execute the original call. Now, what is involved in that? Well, you have to locate the server procedure, and once you locate the server procedure, you have to dispatch the server procedure. And once you dispatch the server procedure, you have to unmarshal the network packet that comes in as the actual arguments for the call that the server procedure has to execute. So all of that setup is first done, and then the server procedure can actually execute the call. So this is the five-step process. From the time the client says, I want to make a RPC call to the point of actually executing the call, these are the layers of software and, of course, hardware and time on the wire, by which time you're ready to execute the server procedure. So even though it looks like a simple procedure call from the clients point of view, there is all this latency to be incurred in executing a remote procedure call. So at the end of step five, the server is all set to execute the procedure. Let's see what happens then. So step number six is server execution, meaning that it is actually executing the procedure. And of course, this is not under our control as operating system designer, because at this point, the amount of time that the server is going to execute this procedure depends really on the logic of the program that has been written as a client server program. And then, finally, once the server procedure has completed execution, then it says okay, I am ready to send the reply back to the client, and that's where we pick up again, so, what happens is that at that point, you are receiving the results. So, once again, just like when the client wanted to send the arguments, you have to convert the actual arguments into a network packet and send it out on the wire. Similarly,when the server is ready to reply, you have to take that reply, which is the results of the execution of this procedure, and make it into a network packet. And at this point, once it has been made into a network packet it has to be handed over to the controller, and the controller does exactly what we did on this side, which is to say the controller latency is gon, is going to be incurred. So that's why you see item number two appearing all over again in the return path. Similar to sending the arguments over to the server on the wire, the results have to be sent on the wire back to the client. And so you see that item number three, which is the time on the wire, is reappearing on the return path as well. Come over to this side. The incoming result message is going to result in a, an interrupt on the receiving node, the client node. And that is exactly similar to what happened on the server side item number four. So you see number item number four reappearing on the return path as well. So that is the interrupt handling part. And once that interrupt is handled, the operating system on the client side said, oh this was on behalf of this client, let me redispatch the client, set up the client so that the client can then receive the results, and restart execution where it left off. So the only two new things that we added in the return path was item number six and seven. Two, three, and four was exactly the same as what we saw on the way over to the server, that is being repeated on the way back to the client. So that's the seven step latency involved in the RPC, not worrying about the actual execution time of the server core itself because that is not in the purview of the operating system, it is in the purview of the client server program that the app developer has done.

Sources of Overhead on RPC

Now that we understand the components of RPC latency, let's understand the sources of overhead that creeps in in carrying out all the different functions, going from the client to the server and back to the client. So far as the client is concerned, this looks like an innocuous procedure call, right? So it just says, I want to call a procedure S.foo, and here are the arguments. Well unfortunately, this call is not a simple procedure call but it is a remote procedure call. And the sources of overhead that creeps in, in a remote procedure call, are marshaling, data copying, control transfer and protocol processing. So we'll look at each one of these things in more detail. Now how can we reduce the overhead in the kernel? What we want to do is, think what the hardware gives you in order to reduce the latency incurred for each of these components of the RPC latency.

Marshaling and Data Copying

First let's look at how we can reduce the overhead and marshaling the arguments and the data copying. Just to jog your memory, marshaling refers to the fact that the semantics of the RPC call being made between the client and the server. It's something that the operating system doesn't have any clue about. So in other words, the arguments that are actually passed between the client and the server, has semantic meaning only between the client and the server. The operating system has no knowledge of it. And therefore marshaling is the term that is used to say, let's accumulate. All of the arguments that is there in the call and make one contiguous network packet out of it, so that we can give it to the kernel and the kernel can send it out. That's what is being described as marshaling, and the biggest source of overhead in marshaling is the data copying that's going to happen, and I'll explain that in a minute. Potentially, in doing the marshaling, there could be three copies involved. Where are these three copies coming aboard? Well, first of all, the client is executing. When a client is executing a procedure, all the arguments for the procedure call that it wants to do is living on the stack of the client. And there is an entity, we'll introduce this terminology even before, called the client stub, and the role of the client stub is to take the arguments of the call which are living on the stack. And convert it into a contiguous sequence of bytes called an RPC message, so the RPC message has no semantic meaning, and it's just a contiguous string of bytes, which you can pass to the kernel, and the kernel can then send it out on the wire, just like any other message. So that's the first thing that the stub does, and that's the first source of overhead. The client stub is making the first copy, from the stack, in order to create an RPC message. Now remember that the client is a user program, so it is living in the user space outside the kernel, and so this RPC message, which is being created by the stub, it is pulled off the client's address space. Which is living outside the kernel. So, this RPC message is in user space and the kernel has to make a copy of the RPC message from the user space into its own buffer, the kernel buffer. And that's a second source of overhead. The second source of copy in doing the marshalling of the arguments. So now it is in the buffer of the operating system kernel, now the operating system can kick the network controller and say, hey, go ahead, take this buffer, send it out in the wire to the desired destination. And the network controller, at that point, is going to move the bits from the buffer, which is in the system memory, of the operating system, into its internal buffer using DMA. And this is the third copy that is happening. The copy that is done by the network controller, using DMA to move the bits of the RPC message, copied from the user space into the internal buffer of the kernel. And now this movement is being orchestrated by the hardware to move it from the kernel buffer into the internal buffer of the network controller, so that it can then get out on the wire. So those are the three copies involved in marshaling the arguments of the call, before it can be put out on the wire. And the copying overhead is the biggest source of overhead for RPC latency. Now, how do we reduce the number of copies? Well, it turns out that the third copy that you're looking at here, moving the bits from the system memory into the network controller, there's a hardware action. That is unavoidable, and therefore, we're going to live with it. Unless the network controller is completely redesigned, if the network controller is saying, well, I need to DMA the bits from the system memory into my buffer. Well, I have to DMA the bits from the system memory into my internal buffer, before I can put it all on the wire. Then this third copy is inevitable so we live with it. But we would like to see if we can try to reduce the number of copies involved here. The first idea is, can we eliminate this copy that is done by the client stub? Why is that happening? Well, it has to create a network message in order to send it out on the wire. As we said. That the semantics of this call is only known to the client and the server, and the client stub is taking the argument and making a network packet out of it. And it was doing it in user space. And so what we're going to do is, we're going to marshal it directly into the kernel buffer. In other words, we've now moved this stub, the client side stub, from the user space down into the kernel. If you can move it into the kernel, then from the stack a stub can directly marshal it into the kernel buffer. And so that intermediate copy that we had here creating an RPC message and copying it. Again into the kernel buffer is avoided if the stub can directly work on the stack, and write it into the kernel memory. So what this means is that, at instantiation time, the client stub is installed inside the kernel. At bind time, when the client binds with the server at the bind time, what we going to do is, we going to say that here is the client stub, please put it inside the kernel so that, later on, you can use that in order to do the marshaling. So the synthesized procedure is installing the kernel for each client call, so for each client server relationship. We synthesize the procedure, which is the client's job, install it in the kernel for use every time, I make this call. This stub can be invoked to convert the argument that are living on the stack, into a network message, and directly put it into the kernel buffer. So this, obviously, will eliminate. From the two copies down to one copy because, the intermediate copy of converting the arguments into an RPC message is now eliminated. Now, the problem with this idea is that we're seeing, let's dump some code into the kernel and that may not be something that is, so palatable. So this is a solution that's possible if the RPC service that is being provided between the client and the server, the trusted service and therefore we can trust, who is putting the stub into the kernel. In that case the solution maybe a reasonable one to adopt.

Marshaling and Data Copying (cont)

An alternative to dumping code into the kernel, is to leave the stub in the users piece itself. But have a structured mechanism, for communication between the client stub and the kernel. And that structured mechanism is a shared descriptor. And the shared descriptor, is a way by a which the sub can describe to the kernel that here is some stuff sitting in the user space. And I am going to tell you how exactly you can extract this information, from the user space and construct it into this buffer for transmission on the wire. Recall what I told you earlier and that is the kernel has no idea, of the semantics of this call and therefore, it has no knowledge of the actual arguments. The data structures that are being used in the call. So what we're going to do is use the shared descriptor as a vehicle for the stub to communicate to the kernel the data structures that need to be passed. So for instance, let's say that the argument for the call has four parameters. Then this descriptor has four entries, and each entry is saying, this is a starting point of a particular Data item, and this is the length of the data item. This is the starting point of the second data item, and this is the length of the data item. Third data item, fourth data item. Kernel doesn't have to know the semantics of these data items. All it needs to know is, what is the starting address for a particular data item, and what is the length of the data item. That's all it needs to know. And this is the descriptor that. Allows the stub to inform the kernel about the arguments, how many arguments there are and what are the size of each argument. It doesn't have to tell the kernel, oh, here is an integer, here's a floating point, here's an array. No, none of that. All that the stub is doing is, it's saying. Here is the starting address for an argument, and here is the length of that argument. Because usually data structures are organized contiguously, so if you have a, an integer, it is occupying full contiguous bytes and memory. If you have floating point number, it is occupying some number of contiguous bytes in memory, and therefore. What the stub is doing is, is creating the shared descriptor that is providing the information of the kernel in the layer of the arguments on the stack, and once the layer of the arguments and the stack are known to the kernel, then the kernel can use these contiguous data items. That are living on the stack, describe the shared descriptor, and create a contiguous packet in its internal buffer. That's a second way you can do, in order to reduce the number of copies from two to one. So in both cases, what we have done is either the first approach of pushing the client stub into the kernel or the second approach of having a shared descriptor between the user stub, which is living. In user space and the kernel in order to describe the layout of the data structures, that need to be assembled into a data packet by the kernel using the shared descriptor. Both of these allow us to reduce the number of copies from three down to two. Either the one copy that is happening Going from this stack into the kernel buffer and this second copy, as I said, is unavailable if the network controller is requiring DMA to be done from the system memory into its internal buffer before the bits can be pushed out of the wire. So that's the first source of overhead, these are techniques that we have looked at. Two differnet techniquest for reducign the copying overhead that is the dominate part of marshalling the arguments. And this happens on both sides. It happens when the client has to push the arguments to the server side. And it happens again on the server side when the server has to push the results back to the client. So the marshalling is happening on both ends, and for both ends we can Use this technique of using a shared descriptor or pushing the clients dub or the service dub into the kernel in order to reduce the number of copies from two down to one.

Control Transfer

So the second source of overhead is the control transfer overhead. And that is the context switches that have to happen in order to effect an RPC call and return. And let's now look at the context switches that happen in doing the RPC. Let's say this is the client machine, and on the client machine the client is making a call. So, they're making the call Another call is made. The kernel has to say, oh, this client wants to make an RPC call and we know that the semantics of RPC is that, the client is blocked until the results are returned. And therefore, the operating system on the client side, will switch from the current client that is making the call, to another process, let's call it C1. So this is the first context which that's going to happen. On the client box, the RPC call is sent out in the wire, reaches the server machine, and when it reaches the server machine, the server machine is executing some arbitrary process let's call it S1, so when the call comes in, the kernel has to switch to the particular server process, that is going to handle this incoming RPC call. So this is the second context switch. So, the server machine and the operating system on the server machine is currently executing some process S1. So it has to S in order to answer the incoming RPC call. So that is the second context switch that happens. Then the server procedure executes. And once the server procedure is completed execution, it's now going to send the results out, and when it wants to send the results out, at that point, the work is done for the server. And so the server operating system has to switch from S to some other process S 2, so that's again a context which that's going to happen because the server is done with whatever it has to do. So that's the third contact switch. Then the RPC result is coming over the wire. Come to the client side. Exactly similar to what happened over here. When it comes back to the client side, the kernel at that point is executing some process, C2. And this particular result message is coming back, saying well the original call sent out on behalf of this client's seed, the result have come back. Now it is time to reschedule this client, so that this client can receive the results, and continue with his execution. Remember that the semantic it's like a procedural call, but it is a remote procedural call, the client is blocked For the result to come back and when the result comes back. The kernel can schedule the client to continue with its execution. So that's what is going on. So potentially, there are four contact switches that are going on. Now let's look at these contact switches and figure out what contact switches are critical. Now, this contact switch is essentially to make sure that, the client box is not being underutilized, right? So once the client has made this call, til the result comes back, the client box is underutalized, and therefore, the operating system says, well let me switch to some other process that can lose some useful work on this node. So that is, this contact switch. Not critical from the point of view of the latency for RPC. Now when the message comes over here. This context switch is crucial because, at this point, when the RPC call comes in, this guy, the server box is executing some of the process S1. So it has to switch to this server process, S, which can actually execute the RPC call. So this is an essential part of the RPC latency. And, similarly, to this context which, that I talked about This contact switch is happening in order to make sure that the server's machine is not underutilized. When the server is done with the RPC call, it's going to send the results back, and therefore, we need the contact switch out of this server process. To some of the process that we can utilize this server box. That's this contact switch. Again, similar to this context switch, this context switch is not in the critical path of RPC latency.

Control Transfer (cont)

The result comes back and when the result comes back the client box is executing some process C2. The kernel has to switch to this client, so that is can see the results, an continue with it's execution. So this context switch, is again in the critical path of RPC latency. So if you look at RPC call, the two contact switches that are in the critcal path of RPC latency. There's a context switch that happens here and the context switch that happens here. So only two context switches are in the critical path of the latency, the context switch that happens on the server machine when the call comes in, and similarity, the context switch that happens on the client machine when the results come in. So this context switch that happens on the client machine to keep the client machine Utilizied, can be overlapped with the network communication to send the call out. So in other words, this context which is not the critical path of RPC latency and therefore do this context which, while the RPC call. Is in transmission on the wire. So while overlapping the context switch that happens on the client box after the call has been sent out with the communication time on the wire for the opposite call. Similarly on the server side, once the server is completed execution and it is ready to send the results out, send it out of the wire and in parallel with sending it out in the wire, it can overlap The contact that happens here in order to keep the server box busy doing useful stuff, S2, S2 that can be overlapped with this network communication. So, only this contact switch and this contact switch are in the critical path of latency. So, we can reduce the number of contact switches down to two. Originally we started with 4, we can reduce it down to 2. By observing that the context which is that happened on the client end box and the server box to keep them utilized can be overlapped with a network communication for sending the arguments over to the server, or sending the results over to the client. Of course we are greedy. Can we reduce it to one? Can we actually reduce the number of context switches Down to one. Let's think about this. So we said that when this RPC call was made, the operating system on the client side said, well, this is a blocking semantic, and therefore, this guy is not going to do any useful work, so I'm going to block him and wait for the results to come in. So this context switch that c, the operating system did on the client side Was essentially to keep this client box from being under-utilized, but do we really need to do the switch? Well, it really depends. If the server procedure is going to execute for a long time, then, you know, this client box is going to be under-utilized for a long time. And in that case, it might be a good thing to context which in order to make sure that we are utilizing the resources that we have. But on the other hand, If, suppose, this RPC call, we know that this RPC call is going to come back very soon. And if it is on a local area network and the server procedure that is going to be executed is not going to take a long time. Then perhaps that RPC call will come back very quickly. If that is the case, we can get rid of this context switch that we talked about here. In order to keep the client box busy, we did this context switch. Don't do that. We can spin instead of switching on the client side. And if you do that, then the client is reading but is not being context restored. It is just that the box is underutilized, so the only context which that we incur. Is the context which on the server because you never know when an RPC call is going to come in. So when an RPC call comes in, you obviously have to contact switch into the server context in order to execute the call. That's the necessary evil. We'll incur that. But on the client side, what we're going to do is, we're going to spin instead of switching so that. Even though the box is underutilized, you're not doing anything on the client side, just sending the call out and waiting. And in that case, we've gotten rid of the second context switch that you need to incur. Because another context was stalled and therefore this context switch which we said is in, inevitable, because. It has to be done in order to receive the results for the client. Well, we can get rid of it if we never switched in the first place, and that's the trick here. To reduce the number of conflicts which is down to one, we can spin on the client side instead of switching so that we can be ready to receive. The results of the RPC call execution when the server is done. Again, the intent here is that we wanted to the latency that is incurred in the RPC call. And since these two contract switches were in the critical path of the latency, we would really like to see how we can elimintate at least one of them. And this context which is inevitable and this context which we can eliminate, by spinning on the client side instead of switching in the first place.

Protocol Processing

So we talked about marshalling, data copying, and content switches, and the fourth component that adds to the latency of the RPC transmission is protocol processing, and that's the next thing that we're going to look at. Now the question is, what transport should we use for the RPC? And this is where we want to see how we can take advantage of what the hardware is giving us. If we are working in a local area network, the local area network is reliable, and therefore our focus should really be on reducing the latency and not worry so much about reliability. It is often the case that performance and reliability are at odds with each other. If you focus on reliability, then performance may take a back seat. So here, since the RPC performance is the most critical thing that we are worried about, we're going to focus on reducing the latency. And we're going to assume that the LAN is reliable and therefore let's not worry about the reliability of message transmission in a local area network. That's the idea behind, the next thing that we're going to look at. Let's think about all the thing that could go wrong in message transmission, and see why some of those things may not be that important, given that we have a reliable local area network. The first thing is, you send a message, it might get lost. But, if in a local area network, the chances that messages will actually get lost, is not very high. It happens in wide area internet, because messages have, have to go out through several different routers, and they maybe queuing in the routers, and there may be loss of packets in the wire and so on. But that's not something that you have to worry about in a local area network. So that assumption that messages may not get lost, suggests that there's no need for low level acknowledgements. Why? Because you're sending a call and the call is going to be executed and the result is going to come back. And usually in network transmission, we send acknowledgements to say that, yes, I received the message. Now, in this case because the semantics of RPC says that the act of receiving the RPC call is going to result in server procedure execution and the result is going to come back, the result itself serves as the ACK. And therefore we don't need low level ACKs to say, oh, I received you arguments of the call. You don't have to do that. And similarly, you don't have to have a low level ACK that says oh, I received the results. Because the results were not received, the caller, the client is going to resend the client call. So the high level semantic of RPC can itself serve as a way we can coordinate between the client and the server and we can eliminate low level ACKs and if we eliminate low level ACKs, that reduces a latency in the transport. The second thing is in message transmission on the Internet, we worry about messages getting corrupted. Not maliciously or anything like that, but just due to vagaries of the network messages may get corrupted in going on the wire that connects the source and destination. And for that reason, it's typical to employ checksum in the messages to indicate the integrity of the message that checksum is usually computed in software and appended to the message and sent on wire. But in a local area network things are reliable, we don't have to do extra overhead and software for generating the checksum, just use hardware checksum if it is available, just use hardware checksum for packet integrity. Don't worry about adding an extra layer of software in the protocol processing for doing software checksum. So that's the second optimization that you can make to make the protocol processing leaner. The third source of overhead that comes about in message transmission is once again related to the fact that messages may get lost in transmission. And therefore in order to make sure that if messages are lost in transmission, you usually buffer the packets. So that if the message is lost in transmission, you can re-transmit the package.

Protocol Processing (cont)

Now once again let's think about the semantics of RPC. The client has made the call and the client is blocked and since the client is blocked, we don't need to buffer the message on the client's side. If the message gets lost for some reason, you don't hear back the result of the RPC from the destination, in a certain amount of time, you can resend the call, from the client side. And therefore you don't have to buffer the client side RPC message but it can reconstruct the client side message and resend the call. Therefore client side buffering is something that you can get rid of, once again, because the LAN is reliable. The next source of overhead, similar to client side buffering, happens on the server side, and that is the server is sending the results and the results may get lost. LAN reliable may not happen that often but it could happen ,and therefore we do want do the buffering. On the server side because if we don't buffer it then you have to reexecute the server procedure to produce a result and that's not something that you want to do because it involves reexecuting the server procedure which may be much more latency intensive then simply buffering the packet that corresponds to the result of executing the server procedure so you do want to buffer on the server site but. The buffering on the server side can be overlapped with the transmission of the message. So in other words the result has been computed by the server procedure. Now go ahead and send the result. While you are sending the result back to the client, do the buffering. That you can overlap the service side buffering with the result transmission, and get it out of the critical path of the latency for protocol processing. So, removing low level asks, employing hardware check sum and not doing check sum in software. Eliminating client side buffering all together. And overlapping the server side buffering with the result transmission are optimizations that you can do. In protocol processing, recognizing that the LAN is reliable and therefore we don't have to focus so much on the reliability of message transmission. But focus rather on the latency, and how we can reduce the latency, by making the protocol processing lean and mean


So once again recapping what we said. The sources of RPC latency are the following. Marshaling and data copying. Context switches, both at the client side and the server side. Similarly marshaling and data copying also happens both in the client side and the server side. And the actual protocol processing in order to send the packet on the wire. These are all the things that are happening in software. Those are the things that as OS designers, we have a chance to do something about. And what we saw were techniques that we can employ for each one of those, reduce the number of copies, reduce the number of context switches, and make the protocol processing lean and mean so that the latency involved in RPC is reduced to as minimum as possible from the software side. And we are going to take whatever the hardware gives us. If the hardware gives us an ability to, to do DMA from the client buffer, we'll use that but if it doesn't, then we have to anchor that. So that's what we are seeing here. As the opportunities for reducing the total latency in going from the client to the server and back to the client in RPC.