Optimization and evaluation of parallel I/O in BIPS3D parallel irregular application Performance Modelling, Evaluation, and optimization of Parallel and.

Optimization and evaluation of parallel I/O in BIPS3D parallel irregular application

Performance Modelling, Evaluation, and optimization of Parallel and Distributed Systems (PMEO-PDS 2007)

Rosa Filgueira, David E. Singh, Florin Isaila, Jesús Carretero, Antonio G. Loureiro

Sumary

I. Description of the problem

II. Main objetives

III. Parallel I/O storage

IV. Evaluation

V. Optimization the I/O

VI. Conclusions

I. Description of the problem (I)

BISP3D is a semiconductor devices simulator based on finite element methods.

Optimization and evaluation of parallel I/O for the BISP3D .

I. Description of the problem (II)

The mesh is divided into several sub-domains (METIS).

Each processor makes calculations only with local data.

The results are stored in a sequential way.

The sequential storage is an important bottleneck.

II. Main objectives (I)

Objetives:

Evaluation of the sequential I/O cost.

Implementing parallel I/O techniques.

Developing a method for selecting the most appropriate I/O technique based on the network type, mesh size and data set size.

Introducing a new data clustering technique called Interval Data Grouping (IDG).

II. Main objectives (II)

Several I/O configurations has been implemented and evaluated: Sequential I/O over NFS. Sequential I/O over PVFS. Parallel I/O over PVFS (unoptimized). Parallel I/O over PVFS with two phase I/O. Parallel I/O over PVFS with List I/O.

III. Parallel I/O All processors write on the disk their local data. Each processor constructs a view over the file using the

distibution provided by METIS.

1 32 4 5 6 7 98 10 1211 1413 15 16 17 18 19 20

1 2 4 98 1211 1413 15

3 5 6 7 10 16 17 19 20

View over the file for processor 1

View over the file for processor 2

1 32 4 5 6 7 98 10 1211 1413 15 16 17 18 19 20

Metis distribution for partition 0 Metis distribution for partition 1

IV. Evaluation (I) We have make tests:

Different networks (Myrinet and Fast Ethernet), Different meshes.

Mesh 1 Mesh 2 Mesh 3 Mesh 4

Nodes 47219 32888 73260 289650

Vertices 305120 210437 416950 2027885

IV. Evaluation (II)

1 32 4 5

1 21’ 2’ 3 3’ 4 54’ 5’

1 1’’1’ 2 2’ 2’’ 3 3’’3’ 4 4’’4’ 5’5 5’’

Mesh with load 2

Mesh

Mesh with load 3

Using a parameter (Load) we increase the size of the mesh

Note that with this parameter we change the grain size of the acceses

IV. Evaluation: Myrinet

Two phase List I/O

IV. Evaluation: : Fast Ethernet

List I/O

IV. Evaluation: Decision tree

Nx=70,000

Nld=50Nld=90

V. Optimizing the I/O

We introduce a novel technique of data grouping: Interval Data Grouping (IDG)

The goal of IDG: grouping data for I/O in order to increase the locality and reduce the disk write time.

V. Optimizing the I/O : Distribution of example mesh

0 21 3 4 5 6 7

0 1 2 3 4

1 2 4 5 6

BISP3D Data distribution

Processor 0

Processor 1

Local

Shared

5 7

7

0 1 3 5 7

2 4 6

Processor 0

Processor 1

METIS assignation

V. Optimizing the I/O : Distribution of example mesh (II)

IDG algorithm has two stages: Node classification:

Analyze the mesh structure and Metis distribution to clasifying mesh node (shared or local):

Disk access scheduler: For local nodes they are written by processor which belong

to For shared nodes we have to choose the most appropriate

one looking its previous and subsequent node.

V. Optimizing the I/O : Distribution of example mesh (III)

0 21 3 4 5 6 7

0 1 2 3 4

6 7

5 Processor 0

Processor 1

Local

Shared

IDG distribution0 1 2 3 4

1 2 4 5 6

BISP3D Data distribution

Processor 0

Processor 1

5 7

7

0 1 3 5 7

2 4 6

Processor 0

Processor 1

METIS assignation

V. Optimizing the I/O : evaluation (I)

We have combined IDG with List I/O for different meshes and different loads.

We have compared the IDG performance with other strategies:

METIS Original node distribution. Random Each shared node is assigned to partition

radomly. First Position Each shared node is assigned to the

first particion among all that it belongs.

V. Optimizing the I/O : evaluation (II)

VI. Conclusions

Optimization and evaluation of parallel I/O operations for BISP3D simulator.

A decision tree to choose the best I/O configuration was made.

We have introduced a novel technique which exploits the data replication of mesh nodes for scheduling disk accesses .With this proposal the perfomance of the parallel I/O operations is improved.