Episode 3. Principles in Network Design

Transcription

1 Episode 3. Principles in Network Design Part 2 Baochun Li Department of Electrical and Computer Engineering University of Toronto

2 Recall: Designing the network as a system Last episode: Every complex computer system involves one or more communication links, usually organized to form a network Identified challenging properties of a network The layering principle: the three-layer reference design The end-to-end argument: applications know the best! 2

3 But there are more of these principles (techniques) in design

4 Reading: Keshav

5 What is system design? A computer network provides computation, storage and transmission resources System design is the art and science of putting together these resources into a harmonious whole Extract the most from what you have 5

6 Performance metrics and resource constraints In any system, some resources are more freely available than others Think about a high-end laptop connected to Internet by a DSL modem The constrained resource is link bandwidth CPU and and memory are unconstrained We wish to maximize a set of performance metrics given a set of resource constraints Explicitly identifying constraints and metrics helps in designing efficient systems e.g., maximize reliability and MPG for a car that costs less than $10,000 to manufacture 6

7 Real-world system design Criteria such as scalability, modularity, extensibility, and elegance are important, but not quantifiable Rapid technological change can add or remove resource constraints an ideal design is future proof Market conditions may dictate changes to design halfway through the process International standards, which themselves change, also impose constraints Nevertheless, still possible to identify some principles 7

8 Most resources are a combination of time, space, computation, money, labor, and scaling

9 Time Shows up in many constraints deadline for task completion time to market mean time between failures Metrics response time: mean time to complete a task throughput: number of tasks completed per unit time degree of parallelism = response time * throughput 20 tasks complete in 10 seconds, and each task takes 3 seconds => degree of parallelism = 3 * 20/10 = 6 9

10 Space Example: a limit on the memory available to hold packets in switches and routers We can also view bandwidth as a space constraint A T3 link has a bandwidth of Mbps. If we use it to carry video streams with a mean bit rate of 1.5 Mbps, we can fit at most 29 streams in the link 10

11 Scaling A design constraint, rather than a resource constraint Minimizes the use of centralized elements in the design forces the use of complicated distributed algorithms Hard to measure but necessary for success 11

12 Common design techniques Key concept: bottleneck the most constrained element in a system System performance improves by removing the bottleneck but creates new bottlenecks In a balanced system, all resources are simultaneously bottlenecked this is optimal, but nearly impossible to achieve in practice, bottlenecks move from one part of the system to another example: Ford Model T 12

13 Top level objective Use unconstrained resources to alleviate bottleneck But how do we do this? Here are several common design techniques that help us to tradeoff one resource for another 13

14 Multiplexing Another word for sharing Trades time and space for money Users see an increased response time, and take up space when waiting, but the system costs less economies of scale make a single large resource cheaper 14

15 Multiplexing Examples multiplexed communication links cloud computing Another way to look at a shared resource unshared virtual resource Server controls access to the shared resource uses a schedule to resolve contention choice of scheduling: critical in proving quality of service guarantees (think about boarding a flight) 15

16 Statistical Multiplexing Suppose resource has capacity C Shared by N identical tasks Each task requires capacity c If Nc <= C, then the resource is underloaded If at most 10% of tasks active, then C >= Nc/10 is enough we have used statistical knowledge of users to reduce system cost this is the statistical multiplexing gain 16

17 Two types of statistical multiplexing Two types: spatial and temporal Spatial we expect only a fraction of tasks to be simultaneously active Temporal we expect a task to be active only part of the time its average resource consumption is less than its peak e.g. silence periods during a voice call; video streams with variable bit rates 17

18 Parallelism: trading computation for time Suppose you wanted to complete a task in less time Could you use more processors to do so? Yes, if you can break up the task into independent subtasks such as downloading images into a browser optimal if all subtasks take the same time What if subtasks are dependent? for instance, a subtask may not begin execution before another ends such as in the iphone assembly line Then, having more processors doesn t always help 18

19 Pipelining Special case of serially dependent subtasks: a subtask depends only on the previous one in execution chain 19

20 Batching: trading response time for throughput Group tasks together to amortize overhead Only works when overhead for N tasks < N time overhead for one task Also, time taken to accumulate a batch shouldn t be too long We re getting reduced overhead and increased throughput, but suffering from a longer worst case response time 20

21 Hierarchy Recursive decomposition of a system into smaller pieces that depend only on their parent for proper execution No single point of control Highly scalable Leaf-to-leaf communication can be expensive: shortcuts help 21

22 Randomization Allows us to break a tie fairly A powerful tool Examples resolving contention in a broadcast medium randomized routing choosing multicast timeouts 22

23 Soft State State: memory in the system that influences future behaviour VCI translation table in ATM networks Problem: needs to create and remove it explicitly The idea of soft state: remove on a timer If you want to keep it, refresh Automatically cleans up after a failure Trades bandwidth and computation for robustness and simpler system design 23

24 Separating data and control planes Divide actions that happen once per connection from actions that happen once per packet Data path: per-packet actions Control path: Actions not in the data path Can increase throughput by minimizing actions in data path Examples connection-oriented (ATM) networks Software-defined networking (SDN) On the other hand, keeping control information in data element has its advantages: more resilient to failures with less state, and per-packet Quality of Service 24

25 Reading: Keshav