Ramana Kompella Lecture 3: Principles CS 636 Internetworking 1 Posted on the web page  http://www.cs.purdue.edu/homes/kompella/teaching/sp09/cs636/restricted/projects.html  If you have not formed a group yet, please do so immediately. CS 636 Internetworking 2 Systems Principles: 1-5 ◦ System is constructed from subsystems ◦ System-wide view versus blackbox  Retain modularity: 6-10 ◦ Improve performance retaining modularity  Speeding it up: 11-15 ◦ Techniques for speeding up individual routines CS 636 Internetworking 3 Amazingly, many of these principles have been used for years by Chef Charlie at Greasy Spoon restaurant! Avoid obvious waste in common sequence of operations  Common sub-expression elimination in compilers ◦ S1: i = (5.1*n + 2) and S2: j=(5.1*n+2)*4  Example in our context: ◦ Copy avoidance.  Chef Charlie: trips to the pantry to get ice-cream maker and pie dish while making pie a la mode CS 636 Internetworking 4 Shift computation in time: ◦ P2a : Pre-compute: Chef Charlie prepares crushed garlic in advance, pre-compute TCP headers ◦ P2b: Evaluate lazily: Dishwashing, copy-on-write in Mach, byte orders in packet data ◦ P2c: Share expenses: Several pies in one oven, timing wheels. CS 636 Internetworking 5 Relax system requirements  P3a: Trading certainty for time: ◦ Ethernet, Approximate statistics counters  P3b: Trading accuracy for time: ◦ MPEG, lossy compression, scheduling algorithms, divides with shifts  P3c: Shifting computation in space: Path MTU discovery CS 636 Internetworking 6 Property P Subsystem 2 Subsystem 1 Property Q Subsystem 2 Subsystem 1 Spec S Weaker Spec W Leverage other system components:  P4a: Exploit local access costs: ◦ Disk algorithms like B-trees ◦ IP lookups using multi-bit tries  P4b: Trade memory for speed: ◦ Lookup tables ◦ Compress to fit in cache  P4c: Exploit hardware features: ◦ Strength reduction in compilers (i*4  i+4) ◦ Fast IP checksum computation CS 636 Internetworking 7 Add hardware to improve performance ◦ Microwave for Charlie  P5a: Memory interleaving and pipelining ◦ IP lookups use multiple banks etc.  P5b: Use wide word parallelism ◦ DRAM page mode, Lucent bit vector scheme  P5c: Combine SRAM and DRAM ◦ Statistics counters CS 636 Internetworking 8 Consider replacing unwieldy general purpose routines with more efficient specialized ones  Database caching schemes: ◦ General purpose use LRU ◦ But, query-processing in a loop require MRU CS 636 Internetworking 9 Question the need for excessive generality: ◦ If restrictions provide big gains, consider living with restrictions  RISC multiplications done on firmware  Fbufs provide specialized virtual memory service that allows efficient copies CS 636 Internetworking 10 Consider alternatives to reference implementations found in specifications as long as it has same results.  Charlie knows that when a recipe asks to cut beans and then cut carrots, he can probably interchange steps without danger.  Upcalls : lower layers can call an upper layer for data or advice, seemingly violating rules of modular design. CS 636 Internetworking 11 Consider passing information (that can optimize performance) between organizational layers while preserving structure  Alto file system: Pointers to next file block (hint) but checked against file block number. CS 636 Internetworking 12 Consider changing protocols to pass information (that can optimize performance)  Like Attn: Ms. Harper on fax messages once you know who to correspond with  Tag switching – where tags are carried in packets to aid fast forwarding. CS 636 Internetworking 13 Optimize the expected case  P11a: Use caches ◦ Paging: Worst case 4 accesses, 0 using caches ◦ Header prediction  Question : How to determine common cases ? CS 636 Internetworking 14 Add or exploit state to increase speed  Use of secondary indices in databases  P12a: Compute incrementally  Strength reduction in loop indexes  IP checksums CS 636 Internetworking 15 Optimize the degrees of freedom  TCAM Puzzle ◦ Maintaining gaps between adjacent levels  Multibit tries ◦ Number of bits at each level can be different CS 636 Internetworking 16 Use special techniques for finite universes  Use bucket sorting, array lookup, bitmaps whenever possible  Page lookup using TLBs (table lookup instead of hashing)  Timing wheels (fixed range using circular array) CS 636 Internetworking 17 Use algorithmic ideas  Consider using efficient data structures that have helped protocol implementations  Charlie’s recipe books have elaborate indices and crosslinks for speedy navigation CS 636 Internetworking 18 Principles not to eliminate but foster creativity.  The set of principles is necessarily incomplete.  These principles are not design oriented but are implementation oriented. CS 636 Internetworking 19Performance problems cannot be solved only through the use of Zen Meditation --- By Jeff Mogul, HP labs. CS 636 Internetworking 20 Reducing page download times ◦ Web pages requested via GET message ◦ Images often dowloaded via separate GET messages  Can’t we inline messages to reduce total latency ? ◦ TCP effects: Acks anyway need to be sent. ◦ Client caching CS 636 Internetworking 21 Speeding up signature-based intrusion detection ◦ String searches common (e.g., perl.exe) ◦ Snort searches for each string separately using Boyer-moore method. ◦ Worst happens when 310 rules are matched (30% overhead observed using gprof)  Can’t we reduce this ? ◦ Idea: to use multiset Boyer-moore that works on multiple string patterns simultaneously. CS 636 Internetworking 22 Nice in theory, but does not work in practice.  Two observed problems: ◦ Multiple string matching is not a bottleneck for traces used. ◦ Cache effects (does not fit well in cache when number of string became > 100)  Redone the design with a small number of multi-set BMs that fit well in cache. CS 636 Internetworking 23CS 636