Using Small Abstractions to Program Large Distributed Systems Douglas Thain University of Notre Dame 11 December 2008

Using Small Abstractions to Program Large Distributed Systems Douglas Thain University of Notre Dame 11 December 2008 (And multicore computers!)

Clusters, clouds, and grids give us access to zillions of CPUs. How do we write programs that can run effectively in large systems?

I have 10,000 iris images acquired in my research lab. I want to reduce each one to a feature space, and then compare all of them to each other. I want to spend my time doing science, not struggling with computers. I have a laptop. I own a few machinesI can buy time from Amazon or TeraGrid. Now What?

How do a I program a CPU? I write the algorithm in a language that I find convenient: C, Fortran, Python, etc… The compiler chooses instructions for the CPU, even if I don’t know assembly. The operating system allocates memory, moves data between disk and memory, and manages the cache. To move to a different CPU, recompile or use a VM, but don’t change the program.

How do I program the grid/cloud? Split the workload into pieces. –How much work to put in a single job? Decide how to move data. –Demand paging, streaming, file transfer? Express the problem in a workflow language or programming environment. –DAG / MPI / Pegasus / Taverna / Swift ? Babysit the problem as it runs. –Worry about disk / network / failures…

How do I program on 128 cores? Split the workload into pieces. –How much work to put in a single thread? Decide how to move data. –Shared memory, message passing, streaming? Express the problem in a workflow language or programming environment. –OpenMP, MPI, PThreads, Cilk, … Babysit the problem as it runs. –Implement application level checkpoints.

Tomorrow’s distributed systems will be clouds of multicore computers. So, we have to solve both problems with a single model.

Observation In a given field of study, a single person may repeat the same of work many times, making slight changes to the data and algorithms. In a given field of study, a single person may repeat the same pattern of work many times, making slight changes to the data and algorithms. Examples everyone knows: –Parameter sweep on a simulation code. –BLAST search across multiple databases. Are there other examples?.

Abstractions for Distributed Computing Abstraction: a declarative specification of the computation and data of a workload. A restricted pattern, not meant to be a general purpose programming language. Uses instead of files. Uses data structures instead of files. Provide users with a. Provide users with a bright path. Regular structure makes it tractable to model and predict performance.

All-Pairs Abstraction AllPairs( set A, set B, function F ) returns matrix M where M[i][j] = F( A[i], B[j] ) for all i,j B1 B2 B3 A1A2A3 FFF A1 An B1 Bn F AllPairs(A,B,F) F FF FF F Moretti, Bulosan, Flynn, Thain, AllPairs: An Abstraction… IPDPS 2008 allpairs A B F.exe

Example Application Goal: Design robust face comparison function. F 0.05 F 0.97

Similarity Matrix Construction F Current Workload: 4000 images 256 KB each 10s per F (five days) Future Workload: images 1MB each 1s per F (three months)

Non-Expert User Using 500 CPUs Try 1: Each F is a batch job. Failure: Dispatch latency >> F runtime. HN CPU FFFF F Try 2: Each row is a batch job. Failure: Too many small ops on FS. HN CPU FFFF F F F F F F F F F F F F F F F F Try 3: Bundle all files into one package. Failure: Everyone loads 1GB at once. HN CPU FFFF F F F F F F F F F F F F F F F F Try 4: User gives up and attempts to solve an easier or smaller problem.

All-Pairs Abstraction AllPairs( set A, set B, function F ) returns matrix M where M[i][j] = F( A[i], B[j] ) for all i,j B1 B2 B3 A1A2A3 FFF A1 An B1 Bn F AllPairs(A,B,F) F FF FF F

How Many CPUs on the Cloud?

What is the right metric?

How to measure in clouds? Speedup? –Seq Runtime / Parallel Runtime Parallel Efficiency? –Speedup / N CPUs? Neither works, because the number of CPUs varies over time and between runs. An Alternative: Cost Efficiency –Work Completed / Resources Consumed –Person-Miles / Gallon –Results / CPU-hours –Results / $$$

All-Pairs Abstraction

Wavefront ( R[x,0], R[0,y], F(x,y,d) ) R[4,2] R[3,2]R[4,3] R[4,4]R[3,4]R[2,4] R[4,0]R[3,0]R[2,0]R[1,0]R[0,0] R[0,1] R[0,2] R[0,3] R[0,4] F x yd F x yd F x yd F x yd F x yd F x yd F F y y x x d d x FF x ydyd

Implementing Wavefront Complete Input Complete Output Distributed Master Multicore Master Multicore Master F F F F Input Output Input Output F F F F

The Performance Problem Dispatch latency really matters: a delay in one holds up all of its children. If we dispatch larger sub-problems: –Concurrency on each node increases. –Distributed concurrency decreases. If we dispatch smaller sub-problems: –Concurrency on each node decreases. –Spend more time waiting for jobs to be dispatched. So, model the system to choose the block size. And, build a fast-dispatch execution system.

Model of 1000x1000 Wavefront

500x500 Wavefront on ~200 CPUs

Wavefront on a 200-CPU Grid

Wavefront on a 32-Core CPU

T2 Classify Abstraction Classify( T, R, N, P, F ) T = testing setR = training set N = # of partitionsF = classifier P T1 T3 F F F T R V1 V2 V3 CV Moretti, Steinhauser, Thain, Chawla, Scaling up Classifiers to Cloud Computers, ICDM 2008.

From Abstractions to a Distributed Language

What Other Abstractions Might Be Useful? Map( set S, F(s) ) Explore( F(x), x: [a….b] ) Minimize( F(x), delta ) Minimax( state s, A(s), B(s) ) Search( state s, F(s), IsTerminal(s) ) Query( properties ) -> set of objects FluidFlow( V[x,y,z], F(v), delta )

How do we connect multiple abstractions together? Need a meta-language, perhaps with its own atomic operations for simple tasks: Need to manage (possibly large) intermediate storage between operations. Need to handle data type conversions between almost-compatible components. Need type reporting and error checking to avoid expensive errors. If abstractions are feasible to model, then it may be feasible to model entire programs.

Connecting Abstractions in BXGrid B1 B2 B3 A1A2A3 FFF F FF FF F Lbrown Lblue Rbrown R S1 S2 S3 eyecolor F F F ROC Curve S = Select( color=“brown” ) B = Transform( S,F ) M = AllPairs( A, B, F ) Bui, Thomas, Kelly, Lyon, Flynn, Thain BXGrid: A Repository and Experimental Abstraction… poster at IEEE eScience 2008

Implementing Abstractions S = Select( color=“brown” ) B = Transform( S,F ) M = AllPairs( A, B, F ) DBMS Relational Database (2x) Active Storage Cluster (16x) CPU Relational Database CPU Condor Pool (500x)

What is the Most Useful ABI? Functions can come in many forms: –Unix Program –C function in source form –Java function in binary form Datasets come in many forms: –Set: list of files, delimited single file, or database query. –Matrix: sparse elem list, or binary layout Our current implementations require a particular form. With a carefully stated ABI, abstractions could work with many different user communities.

What is the type system? Files have an obvious technical type: –JPG, BMP, TXT, PTS, … But they also have a logical type: –JPG: Face, Iris, Fingerprint etc. –(This comes out of the BXGrid repository.) The meta-language can easily perform automatic conversions between technical types, and between some logical types: –JPG/Face -> BMP/Face via ImageMagick –JPG/Iris -> BIN/IrisCode via ComputeIrisCode –JPG/Face -> JPG/Iris is not allowed.

Abstractions Redux Mapping general-purpose programs to arbitrary distributed/multicore systems is algorithmically complex and full of pitfalls. But, mapping a single abstraction is a tractable problem that can be optimized, modeled, and re-used. Can we combine multiple abstractions together to achieve both expressive power and tractable performance?

Acknowledgments Cooperative Computing Lab – Grad Students –Chris Moretti –Hoang Bui –Karsten Steinhauser –Li Yu –Michael Albrecht Faculty: –Patrick Flynn –Nitesh Chawla –Kenneth Judd –Scott Emrich NSF Grants CCF , CNS Undergrads –Mike Kelly –Rory Carmichael –Mark Pasquier –Christopher Lyon –Jared Bulosan