Networking Applications Janani Ravi Metha Jeeradit John Kim Outline Introduction Techniques to improve performance Caching Parallelism Example media streaming Network processors Benchmarks 2 Introduction Network can be considered as a distributed information system providing access to shared data Due to rapid growth of Internet networking applications usually span large geographical area raising several issues e g Network latency tolerance Error and loss handling Delay requirements for real time apps 3 Web Caching Proxy servers can be used to cache recently used documents Reduces bandwidth consumption access latencies workload of remote servers and improves availability of cached documents Possible access to stale data extra processing due to cache miss single proxy would be a bottleneck 4 Caching Architectures Hierarchical Shorter access times but Every level adds delay higher level caches bottlenecks multiple copies at different levels Distributed Shorter transmission times and better bandwidth usage but Higher bandwidth usage overall Hybrid 5 Latencies Connection time Transmission time Total latency 6 Prefetching Local Based on local information like reference patterns Server hint Based on content specific knowledge of objects and reference patterns of a far greater number of clients 7 Parallelism Layer level Connection level Functional level Packet Level 8 Example Media Streaming multiple senders one receiver Requirements Ability to tolerate errors and jitters High network bandwidth Delay and performance guarantees 9 Example Media Streaming Receiver side Estimate sender s BW at a regular interval and perform rate allocation TLP in BW estimation since each sender s BW can be estimated independently Sequential computation in rate allocation 10 Example Media Streaming Sender side Keep track of difference between est arrival time and playback time of each packet If max difference for current sender then send packet TLP and PLP in difference estimation PLP in sending the packets 11 Example Media Streaming Computational Complexity Roughly constant with increasing number of senders with enough hardware support Performance Bottleneck Inter process communication between the application and the operating system High context switching cost e g on Windows NT 12 Architecture GPU or NPU History of processors used in networking software based flexibility ASIC performance specialized network processors Programmability Performance Management Routing 13 Architecture GPU or NPU Moore s Law vs Network requirement Performance avg case vs worst case Locality speculation Real time requirement Power budget ILP vs TLP packet level parallelism Multithreaded architecture 14 Network Processors Specialized chips to handle data trafficking 4 network processors evaluated Superscalar Fine grained multithreaded Chip multiprocessor Simultaneous multithreaded 15 Performance Comparison P Crowley et Al May 00 16 Benchmarks Available EEMBC industry standard benchmark in embedded application MiBench cheap version of EEMBC Mediabench multimedia communication Netbench target network processors Commbench packet based processing tasks such as routing data forwarding 17 Benchmarks Other benchmarks 18