12
InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz campus

InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

Embed Size (px)

Citation preview

Page 1: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell foundations with ParXXL

Render Large Scale Computations Interactive

Jens Gustedt

INRIA Nancy – Grand Est

AlGorille

Stéphane Vialle

Supélec Metz campus

Page 2: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 2

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

Goals

Provide a programming paradigm and an interface:Simplicity:

easy to use close to existing programming habits

Performance: competitive responsive easy to evaluate

Interoperability: close to known standards, heterogenity of computing material, OS, administrative domains

Page 3: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 3

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

Page 4: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 4

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

Round trip time:> 0.13 s

SynchronizationFrequency:< 7.5 Hz

Problem 1: Communication Latency

Page 5: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 5

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

Overruling Latency

Page 6: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 6

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

Problem 2: Data Management

Access?Access?

Management?Management?

Page 7: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 7

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

Delegation of responsibilities

Page 8: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 8

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

Isolation in welldefined partsControl Communication

Page 9: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 9

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

Two commonly used paradigms

Message Passing

• Standardization (MPI)

• Efficiency (distributed env.)• Simplified Data Control

• Memory Blowup• Extra Copies

Shared Memory

• Standardization (pthread, shm_open)

• Latency Problems• Data Consistency Problems

• Efficiency (parallel machines)• Random Access during Computation

Page 10: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 10

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

parXXL unifies these paradigms

Low Level• Computational Efficiency (distributed and / or parallel)• Simplified Data Control• Memory Efficiency (data size)• Transfer Efficiency

High Level• Abstraction for Cellular Computation• Implementation of commonly used Cell-Networks

Page 11: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 11

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

parXXL achievements 2008

Low Level• Packaging and Portability• Code Stability:

• Size Up (fill whole InterCell cluster)• Robustness (long runs = transient failures)

• Transfer Efficiency:• Data collectors. Collect distinguished data from the cells• Mix communication models (shared memory, files & messages)

Page 12: InterCell foundations with ParXXL Render Large Scale Computations Interactive Jens Gustedt INRIA Nancy – Grand Est AlGorille Stéphane Vialle Supélec Metz

InterCell lowlevel ParXXL 12

Jens Gustedt INRIA

RILLEpour la g

rithmes ALGO

Stéphane Vialle

ParXXL achievements 2008, cont.

High Level• Cellular Computation:

• Find a compromise between synchronized and de-synchronized Cell-Computation

• Develop new modes of asynchronous computation (cont.)• Implementation of commonly used Cell-Networks:

• Optimize the parallel deployment (avoid scaling problems) (cont.)

... changed 40000 lines of code (sep 07 – sep 08)