Distributed Systems Laboratory Computational Biology Laboratory Superlink-Online: Harnessing the world’s computers to hunt for disease-provoking genes Mark Silberstein, CS, Technion Dan Geiger, Computational Biology Lab Assaf Schuster,
Download ReportTranscript Distributed Systems Laboratory Computational Biology Laboratory Superlink-Online: Harnessing the world’s computers to hunt for disease-provoking genes Mark Silberstein, CS, Technion Dan Geiger, Computational Biology Lab Assaf Schuster,
Distributed Systems Laboratory Computational Biology Laboratory Superlink-Online: Harnessing the world’s computers to hunt for disease-provoking genes Mark Silberstein, CS, Technion Dan Geiger, Computational Biology Lab Assaf Schuster, Distributed Systems Lab Genetics Research Institutes in Israel, EU, US MS eScience Workshop 2008 1 Familial Onychodysplasia and dysplasia of distal phalanges (ODP) III-15 IV-10 IV-7 5 Family Pedigree MS eScience Workshop 2008 6 Marker Information Added Id, dad, mom, sex, aff Marker 1 Marker 2 III-21 II-10 II-11 f h 0 0 0 0 II-5 I-3 I-4 f h 155 157 A A III-7 II-4 II-5 f a 155 157 A T III-13 II-4 II-5 m a 151 155 A T III-14 II-1 II-2 f h 151 155 A A III-15 II-4 II-5 male a 151 155 A A III-16 II-10 II-11 f h 151 159 A A III-5 II-4 f h 151 155 A A IV-1 III-13 III-14 f h 151 155 A T IV-2 III-13 III-14 f a 151 155 A T 155 155 A T IV-3 III-13 II-5 III-14 female a M1 . M2 Chromosome pair: MS eScience Workshop 2008 7 Maximum Likelihood Evaluation M1 M2 D1 M3 M4 202,209 202,202 a h 139,141 139,146 1,2 3,3 D2 θ III-15 151,159 III-16 151,155 The computational problem: find a value of θ maximizing Pr(data|θ) LOD score (to quantify how confident we are): Z(θ)=log10[Pr(data|θ) / Pr(data|θ=½)]. MS eScience Workshop 2008 8 Results of Multipoint Analysis Position in centi-Morgans Ln(Likelihood) LOD 0.0000 (Marker 3) -216.0217 -14.74 0.5500 -192.2385 -4.41 1.1000 (Marker 4) -216.0210 -14.74 3.6000 -176.3810 2.47 6.1000 (Marker 5) -174.3392 3.35 8.6500 -173.9743 3.51 11.2000 (Marker 6) -173.7030 3.63 16.5500 -173.3106 3.80 21.9000 (Marker 9) -172.9497 3.96 25.2500 -173.6540 3.65 28.6000 (Marker 10) -177.5622 1.95 40.3001 -178.9946 1.33 MS eScience Workshop 2008 9 The Bayesian network model Locus 1 Locus 2 (Disease) Locus 3 Locus 4 This model depicts the qualitative relations between the variables. We need also to specify the joint distribution MS eScience Workshop 2008 over these variables.10 The Computational Task ComputingExponential Pr(data|θ)time forand a specific value of θ : space in: n P(data• #variables | ) P ( xi | pai ) five per person xk x3 x1 i 1 #markers #gene loci Finding the best order is equivalent to finding #values per variable the best order for sum-product operations for #alleles non-typed persons high dimensional matrices : table dimensionality cycles in pedigree Yij m A ikl n l BkjmClmn k MS eScience Workshop 2008 11 Divisible Tasks through Variable Conditioning non trivial parallelization overhead MS eScience Workshop 2008 13 Terminology • Basic unit of execution – batch job – Non-interactive mode: “enqueue – wait – execute – return” – Self-contained execution sandbox • A linkage analysis request - a task – A bag (of millions) of jobs – Turnaround time is important MS eScience Workshop 2008 15 Requirements • The system must be geneticists-friendly – Interactive experience • Low response time for short tasks • Prompt user feedback – Simple, secure, reliable, stable, overloadresistant, concurrent tasks, multiple users... – Fast computation of previously infeasible long tasks via parallel execution • Harness all available resources: grids, clouds, clusters • Use them efficiently! MS eScience Workshop 2008 16 Grids or Clouds? Remaining Jobs in Queue Preempted jobs, UW Madison Grid (k CPUs) Long tail due to failures Cloud (k CPUs) Error rate, UW Madison Queuing time in EGEE Time Queue Waiting Time Small tasks are severely slow on grids Takes 5 minutes on 10-nodes dedicated cluster May take several hours on a grid Should we move scientific loads on the cloud? YES! MS eScience Workshop 2008 17 Grids or Clouds? Consider 3.2x106 jobs, ~40 min each It took 21 days on ~6000-8000 CPUs It would cost about $10K on Amazon’s EC2 Should we move scientific loads on the cloud? NO! ? MS eScience Workshop 2008 18 Clouds or Grids? Clouds and Grids! Low Low Opportunistic Reliability Performance predictibility High High Dedicated Potential amount of available resources Low High Reuse of existing infrastructure Low High MS eScience Workshop 2008 19 Cheap and Expensive Resources Task sensitivity to QoS differ in different stages Remaining jobs in queue High throughput High performance Use cheap unreliable resources Grids Community grids Non-dedicated clusters Use expensive reliable resources Dedicated clusters Clouds Dynamically determine entering tail mode Switch to expensive resources (gracefully) MS eScience Workshop 2008 20 Glue pools together via overlay Scheduling Server Job queue Scheduler Issues: granularity, load balancing, firewalls, failed resources, scheduler scalability… Submitter to Grid 1 Submitter to Grid 2 Virtual cluster maintainer Submitter to Cloud 1 Submitter to Cloud 2 21 Practical considerations Overlay scalability and firewall penetration Compatibility with community grids The server is based on BOINC Agents are upgraded BOINC clients Elimination of failed resources from scheduling Server may not initiate connect to the agent Performance statistics is analyzed Resource allocation depending on the task state Dynamic policy update via Condor classad mechanism MS eScience Workshop 2008 22 SUPERLINK@TECHNION Upgraded BOINC Server HTTP frontend Scheduler Database jobs, monitoring, system statistics Web Portal Task state Task execution and monitoring workflow BOINC clients submitter for EGEE BOINC clients submitter for Madison pool Virtual cluster maintainer Submitter to Technion Submitter To EC2 Cloud Submitter to OSG Submitter to any grid/cluster/cloud Dedicated cluster fallback 23 Superlink-online 1.0: http://bioinfo.cs.technion.ac.il 24 Task Submission 25 Superlink-online statistics ~1720 CPU years for ~18,000 tasks during 20062008 (counting) ~37 citations (several mutations found) Over 250 (counting) users: Israeli and international Examples: Ichthyosis,"uncomplicated" hereditary spastic paraplegia (1-9 people per 100,000) Soroka H., Be'er Sheva, Galil Ma'aravi H., Nahariya, Rabin H., Petah Tikva, Rambam H., Haifa, Beney Tzion H., Haifa, Sha'arey Tzedek H., Jerusalem, Hadassa H., Jerusalem, Afula H. NIH, Universities and research centers in US, France, Germany, UK, Italy, Austria, Spain, Taiwan, Australia, and others... Task example 250 days on single computer - 7 hours on 300-700 computers Short tasks: few seconds even during severe overload MS eScience Workshop 2008 26 Using our system in Israeli Hospitals Rabin Hospital, by Motti Shochat’s group New locus for mental retardation Infantile bilateral striatal necrosis Soroka Hospital, by Ohad Birk’s group Lethal congenital contractural syndrome Congenital cataract Rambam Hospital, by Eli Shprecher’s group Congenital recessive ichthyosis CEDNIK syndrome Galil Ma’aravi Hospital, by Tzipi Falik’s group Familial Onychodysplasia and dysplasia Familial juvenile hypertrophy MS eScience Workshop 2008 27 Utilizing Community Computing ~3.4 TFLOPs, ~3000 users, from 75 countries 28 Superlink-online V2(beta) deployment Technion Condor pools Submission server EGEE-II BIOMED VO UW in Madison Condor pool Dedicated cluster ~12,000 hosts operational during the last month Superlink@Campus Superlink@Technion OSG GLOW VO MS eScience Workshop 2008 29 3.1 million jobs in 21 days 60 dedicated CPUs only MS eScience Workshop 2008 30 Conclusions Our system integrates clusters, grids, clouds, community grids, etc. Geneticist friendly Minimizes use of expensive resources while providing QoS for tasks Generic mechanism for scheduling policy Can dynamically reroute jobs from one pool to another according to a given optimization function (budget, energy, etc.) MS eScience Workshop 2008 31 NVIDIA Compute Unified Device Architecture (CUDA) GPU 16MPX8SPX4 MP S P S P S P S P S P S P S P S P S P S P S P S P Cached Read-Only memory S P S P S P S P ... S P S P S P S P S P S P S P S P Cached Read-Only memory Global Memory 33 S P S P S P S P Register file S P S P S P S P Shared memory (16KB) Register file ~1 cycle ~TB/s MP Shared memory (16KB) MS eScience Workshop 2008 Key ideas (Joint work with John Owens -UC Davis) Software-managed cache We implement the cache replacement policy in software Maximization of data reuse Better compute/memory access ratio A simple model for performance bounds Yes, we are (optimal) Use special function units for hardwareassisted execution 34 MS eScience Workshop 2008 Results summary Experiment setup CPU: single core Intel Core 2 2.4GHz, 4MB L2 GPU: NVIDIA G80 (GTX8800), 750MB GDDR4, 128 SP, 16K mem / 512 threads Only kernel runtime included (no memory transfers, no CPU setup time) 2500~ 2 x 25 x 25 x 2 Hardware 35 Software managed Caching Use of SFU: expf is about 6x slower than “+” on GPU, but ~200x slower on CPU Acknowledgments Superlink-online team: Alumni: Anna Tzemach, Julia Stolin, Nikolay Dovgolevsky, Maayan Fishelson, Hadar Grubman, Ophir Etzion Current: Artyom Sharov, Oren Shtark Prof. Miron Livny (Condor pool UW Madison, OSG) EGEE BIOMED VO and OSG GLOW VO Microsoft TCI program, NIH grant, SciDAC Institute for ultrascale visualization If your grid is underutilized – let us know! Visit us at: http://bioinfo.cs.technion.ac.il/superlink-online Superlink@TECHNION project home page: http://cbl-boinc-server2.cs.technion.ac.il/superlinkattechnion MS eScience Workshop 2008 36 QUESTIONS??? Visit us at: http://bioinfo.cs.technion.ac.il/superlink-online MS eScience Workshop 2008 37