1 Teleport Messaging for Distributed Stream Programs William Thies, Michal Karczmarek, Janis Sermulins, Rodric Rabbah and Saman Amarasinghe Massachusetts Institute of Technology PPoPP 2005 http://cag.lcs.mit.edu/streamit Please note: This presentation was updated in September 2006 to simplify the timing of upstream messages. The corresponding update of the paper is available at http://cag.csail.mit.edu/commit/papers/05/thies-ppopp05.pdf
Streaming Application Domain 2 AtoD • Based on a stream of data – Radar tracking, microphone arrays, Decode HDTV editing, cell phone base stations duplicate – Graphics, multimedia, software radio • Properties of stream programs LPF 1 LPF 2 LPF 3 – Regular and repeating computation – Parallel, independent actors HPF 1 HPF 2 HPF 3 with explicit communication roundrobin – Data items have short lifetimes Amenable to aggressive Encode compiler optimization [ASPLOS ’02, PLDI ’03, LCTES’03, LCTES ’05] Transmit
Control Messages 3 AtoD • Occasionally, low-bandwidth control messages are sent between actors Decode • Often demands precise timing duplicate – Communications: adjust protocol, amplification, compression LPF 1 LPF 2 LPF 3 – Network router: cancel invalid packet – Adaptive beamformer: track a target HPF 1 HPF 2 HPF 3 – Respond to user input, runtime errors – Frequency hopping radio roundrobin What is the right programming model? Encode How to implement efficiently? Transmit
Supporting Control Messages 4 • Option 1: Synchronous method call PRO: - delivery transparent to user CON: - timing is unclear - limits parallelism • Option 2: Embed message in stream PRO: - message arrives with data CON: - complicates filter code - complicates stream graph - runtime overhead
Teleport Messaging 5 • Looks like method call, but timed relative to data in the stream TargetFilter x; if newProtocol(p) { x.setProtocol(p) @ 2; } void setProtocol(int p) { reconfig(p); } • PRO: – simple and precise for user • adjustable latency • can send upstream or downstream – exposes dependences to compiler
Outline 6 • StreamIt • Teleport Messaging • Case Study • Related Work and Conclusion
Outline 7 • StreamIt • Teleport Messaging • Case Study • Related Work and Conclusion
Model of Computation 8 • Synchronous Dataflow [Lee 92] A/D – Graph of autonomous filters – Communicate via FIFO channels Band Pass – Static I/O rates Duplicate • Compiler decides on an order of execution (schedule) Detect Detect Detect Detect – Many legal schedules LED LED LED LED
Example StreamIt Filter 9 float-> float filter LowPassFilter (int N, float[N] weights) { work peek N push 1 pop 1 { float result = 0; for (int i= 0; i< weights.length; i+ + ) { result + = weights[i] * peek (i); } N push (result); pop (); } filter }
Example StreamIt Filter 10 float-> float filter LowPassFilter (int N, float[N] weights) { work peek N push 1 pop 1 { float result = 0; for (int i= 0; i< weights.length; i+ + ) { N result + = weights[i] * peek (i); } push (result); pop (); } filter handler setWeights(float[N] _weights) { weights = _weights; } }
Example StreamIt Filter 11 float-> float filter LowPassFilter (int N, float[N] weights , Frontend f ) { work peek N push 1 pop 1 { float result = 0; for (int i= 0; i< weights.length; i+ + ) { N result + = weights[i] * peek (i); } if (result = = 0) { f.increaseGain() @ [2:5]; } filter push (result); pop (); } handler setWeights(float[N] _weights) { weights = _weights; } }
StreamIt Language Overview 12 • StreamIt is a novel filter language for streaming pipeline – Exposes parallelism and may be communication any StreamIt language construct – Architecture independent splitjoin – Modular and composable parallel computation • Simple structures composed to creates complex graphs – Malleable splitter joiner • Change program behavior with small modifications feedback loop splitter joiner
Outline 13 • StreamIt • Teleport Messaging • Case Study • Related Work and Conclusion
Providing a Common Timeframe 14 • Control messages need precise timing with respect to data stream • However, there is no global clock in distributed systems – Filters execute independently, whenever input is available • Idea: define message timing with respect to data dependences – Must be robust to multiple datarates – Must be robust to splitting, joining
Stream Dependence Function (SDEP) 15 • Describes data dependences between filters A B
Stream Dependence Function (SDEP) 16 • Describes data dependences between filters A B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 17 • Describes data dependences between filters A n SDEP A � B ( n ) push 2 0 1 pop 3 2 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 18 • Describes data dependences between filters A n SDEP A � B ( n ) push 2 0 0 1 pop 3 2 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 19 • Describes data dependences between filters A × 1 n SDEP A � B ( n ) push 2 0 0 1 pop 3 2 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 20 • Describes data dependences between filters A × 2 n SDEP A � B ( n ) push 2 0 0 1 pop 3 2 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 21 • Describes data dependences between filters A × 2 n SDEP A � B ( n ) push 2 0 0 1 pop 3 × 1 2 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 22 • Describes data dependences between filters A × 2 n SDEP A � B ( n ) push 2 0 0 1 2 pop 3 × 1 2 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 23 • Describes data dependences between filters A × 3 n SDEP A � B ( n ) push 2 0 0 1 2 pop 3 × 1 2 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 24 • Describes data dependences between filters A × 3 n SDEP A � B ( n ) push 2 0 0 1 2 pop 3 × 2 2 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 25 • Describes data dependences between filters A × 3 n SDEP A � B ( n ) push 2 0 0 1 2 pop 3 × 2 2 3 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Stream Dependence Function (SDEP) 26 • Describes data dependences between filters A × 3 n * 3 = n SDEP A � B ( n ) push 2 2 0 0 1 2 pop 3 × 2 2 3 B SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Calculating SDEP: General Case 27 A SDEP A � C (n) = max [SDEP A � Bi (SDEP Bi � C (n))] B 1 B m i ∈ [1,m] SDEP is compositional C SDEP A � B ( n ): minimum number of times that A must execute to make it possible for B to execute n times
Teleport Messaging using SDEP 28 • SDEP provides precise semantics for message timing S If S sends message to R : • on the n th execution of S • with latency range [ k 1 , k 2 ] X Then message is delivered to R : • on any iteration m such that n + k 1 · SDEP S � R ( m ) · n + k 2 R
Teleport Messaging using SDEP 29 • SDEP provides precise semantics for message timing S push 1 If S sends message to R : • on the n th execution of S pop 1 • with latency range [ k 1 , k 2 ] X push 1 Then message is delivered to R : • on any iteration m such that pop 1 n + k 1 · SDEP S � R ( m ) · n + k 2 R
Teleport Messaging using SDEP 30 • SDEP provides precise semantics for message timing × 1 S push 1 If S sends message to R : • on the n th execution of S pop 1 • with latency range [ k 1 , k 2 ] X push 1 Then message is delivered to R : • on any iteration m such that pop 1 n + k 1 · SDEP S � R ( m ) · n + k 2 R
Teleport Messaging using SDEP 31 • SDEP provides precise semantics for message timing × 2 S push 1 If S sends message to R : • on the n th execution of S pop 1 • with latency range [ k 1 , k 2 ] X push 1 Then message is delivered to R : • on any iteration m such that pop 1 n + k 1 · SDEP S � R ( m ) · n + k 2 R
Teleport Messaging using SDEP 32 • SDEP provides precise semantics for message timing × 3 S push 1 If S sends message to R : • on the n th execution of S pop 1 • with latency range [ k 1 , k 2 ] X push 1 Then message is delivered to R : • on any iteration m such that pop 1 n + k 1 · SDEP S � R ( m ) · n + k 2 R
Recommend
More recommend
