peters HTC BlueGene CondorWeek

High Throughput Computingon Blue Gene: 

High Throughput Computing on Blue Gene IBM Rochester: Amanda Peters, Tom Budnik With contributions from: IBM Rochester: Mike Mundy, Greg Stewart, Pat McCarthy IBM Watson Research: Alan King, Jim Sexton UW-Madison Condor: Greg Thain, Miron Livny, Todd Tannenbaum

Agenda: 

Agenda Blue Gene Architecture Overview High Throughput Computing (HTC) on Blue Gene Condor and IBM Blue Gene Collaboration Exploratory Application Case Studies for Blue Gene HTC Questions and Web resource links

Blue Gene/L Overview: 

2.8/5.6 GF/s 2 processors 2 chips 5.6/11.2 GF/s 1.0 GB 32 chips 16 compute, 0-2 IO cards 90/180 GF/s 16 GB 32 node cards 1,024 chips 2.8/5.6 TF/s 512 GB 64 Racks 65,536 chips 180/360 TF/s 32 TB Rack System Node card Compute node Chip Blue Gene/L Overview Scalable from 1 rack to 64 racks Rack has 2048 processors with 512 MB or 1 GB DRAM/node Blue Gene has 5 independent networks (Torus, Collective, Control (JTAG), Global barrier, and Functional 1 Gb Ethernet)

Blue Gene System Architecture : 

Blue Gene System Architecture Functional Gigabit Ethernet I/O Node 0 Linux ciod I/O Node 1023 Linux ciod Control Gigabit Ethernet IDo chip Resource Scheduler System Console Control System DB2 I2C fs client fs client app app app app

HPC vs. HTC Comparison: 

HPC vs. HTC Comparison High Performance Computing (HPC) Model Parallel, tightly coupled applications Single Instruction, Multiple Data (SIMD) architecture Programming model: typically MPI Apps need tremendous amount of computational power over short time period High Throughput Computing (HTC) Model Large number of independent tasks Multiple Instruction, Multiple Data (MIMD) architecture Programming model: non-MPI Apps need large amount of computational power over long time period Traditionally run on large clusters HTC and HPC modes co-exist on Blue Gene Determined when resource pool (partition) is allocated

Why Blue Gene for HTC?: 

Why Blue Gene for HTC? High processing capacity with minimal floor space High compute node density – 2,048 processors in one Blue Gene rack Scalability from 1 to 64 racks (2,048 to 131,072 processors) Resource consolidation Multiple HTC and HPC workloads on a single system Optimal use of compute resources Low power consumption #1 on Green500 list @ 112 MFlops/Watt (www.green500.org/CurrentLists.html) Twice the performance per watt of a high frequency microprocessor Low cooling requirements enable extreme scale-up Centralized system management Blue Gene Navigator

Slide7: 

Generic HTC Flow on Blue Gene : 

Generic HTC Flow on Blue Gene One or more dispatcher programs are started on front end/service node Dispatcher will manage HTC work request queue A pool (partition) of compute nodes is booted on Blue Gene Every compute node has a launcher program started on it that connects back to the designated HTC dispatcher New pools of resources can be added dynamically as workload increases External work requests are routed to HTC dispatcher queue Single or multiple work requests from each source HTC dispatcher finds available HTC client and forwards the work request HTC client runs executable on compute node A launcher program on each compute node handles work request sent to it by the dispatcher. When work request completes, the launcher program is reloaded and client is ready to handle another work request. Executable exit status is reported back to dispatcher

Generic HTC Flow on Blue Gene: 

Generic HTC Flow on Blue Gene

Node Resiliency for HTC: 

Node Resiliency for HTC In HPC mode a single failing node in a partition (pool of compute nodes) causes termination of all nodes in the partition Expected behavior for parallel MPI type apps, but unacceptable for HTC apps HTC mode partition handles this situation In HTC mode Blue Gene can recover from soft node failures For example parity errors If failure is not related to network hardware, a software reboot will recover the node Other nodes in the partition are unaffected and continue to run jobs Job on failed node is terminated and must be resubmitted by dispatcher If the partition is started in HTC mode, the Control System will poll at regular intervals looking for nodes in the reset state Nodes in the reset state will be rebooted and launcher restarted on them

Condor and IBM Blue Gene Collaboration: 

Condor and IBM Blue Gene Collaboration Both IBM and Condor teams engaged in adapting code to bring Condor and Blue Gene technologies together Initial Collaboration (Blue Gene/L) Prototype/research Condor running HTC workloads on Blue Gene/L Condor developed dispatcher/launcher running HTC jobs Prototype work for Condor being performed on Rochester On-Demand Center Blue Gene system Mid-term Collaboration (Blue Gene/L) Condor supports HPC workloads along with HTC workloads on Blue Gene/L Long-term Collaboration (Next Generation Blue Gene) I/O Node exploitation with Condor Partner in design of HTC services for Next Generation Blue Gene Standardized launcher, boot/allocation services, job submission/tracking via database, etc. Study ways to automatically switch between HTC/HPC workloads on a partition Data persistence (persisting data in memory across executables) Data affinity scheduling Petascale environment issues

Condor Architecture: 

Execute Machine Submit Machine Condor Architecture Submit Schedd Starter Shadow Startd Central Manager Collector Negotiator

Condor with Blue Gene/L: 

Blue Gene I/O Node Submit Machine Condor with Blue Gene/L Submit Schedd Starter Shadow Startd Central Manager Collector Negotiator mpirun Blue Gene Compute Nodes etc.

Exploratory Application Case Studies for Blue Gene HTC: 

Exploratory Application Case Studies for Blue Gene HTC Case Study #1: Financial overnight risk calculation for trading portfolio Large number of calculations to be completed by market opening Algorithm is Monte Carlo simulation Easy to distribute and robust to resource failure (fewer simulations just gives less accurate result) Grid middleware bundles tasks into relatively long-running jobs (45 minutes) Limiting resource is number of CPUs In some cases power density (KW/sq foot) is critical Case Study #2: Molecular docking code for virtual drug screening Docking simulation algorithm for screening large databases of potential drugs against targets Large number of independent calculations to determine the minimization energy between the target and each potential candidate, and subsequently find the strongest leads

Exploratory Application Case Studies for Blue Gene HTC: 

Exploratory Application Case Studies for Blue Gene HTC Experience results: Demonstrated scalable task dispatch to 1000’s of processors Successfully verified multiple dispatcher architecture Discovered optimal ratio of dispatcher to partition (pool) size is 1:64 or less Latencies increase as ratio increases above this level, possibly due to launcher contention for socket resource as scaling increases – still investigating in this area May depend on task duration and arrival rates Running in HTC mode changes the I/O patterns Typical MPI programs read and write to the file system with small buffer sizes HTC requires loading the full executable into memory and sending it to compute node Launcher is cached on IO Node but not the executable Experiments with delaying dispatch proportional to executable size for effective task distribution across partitions were successful Due to IO Node to Compute Node bandwidth To achieve the fastest throughput a low compute node to I/O node ratio is desirable

Questions?: 

Questions? http://www.ibm.com/servers/deepcomputing/bluegene.html http://www.research.ibm.com/bluegene http://www.redbooks.ibm.com/cgi-bin/searchsite.cgi?query=blue+gene Web resources:

Backup Slides: 

Backup Slides

Blue Gene Software Stack : 

Blue Gene Software Stack

Slide19: 

Slide20: 

Dispatcher Launcher Connect to Dispatcher Dispatch task N Start task N Reboot Launcher Connect to Dispatcher andamp; send task N status Exit task N Boot Launcher Write task N status Read task N Submitter Submit task N to Work Queue Read task N status off Results Queue

Slide21: 

Node Resiliency