Typed MPI - A Data Type Tool for MPI

Typed MPI - A Data Type Tool for MPI

Nitin Bahadur

and

Florentina Irina Popovici{ bnitin, pif } @cs.wisc.edu

Nitin Bahadur, Florentina Irina Popovici, Dec 1999

Why a Data Type Tool for MPI?

No type message at receiver end User has to manually construct MPI data types corresponding to user defined

data types

struct Partstruct { int class; double d[6]; char b[7]; };struct Partstruct particle;/* build datatype describing first array entry */MPI_Datatype Particletype;MPI_Datatype type[3] = {MPI_INT, MPI_DOUBLE, MPI_CHAR};int block[3] = {1, 6, 7};MPI_Aint disp[3];MPI_Address( particle, disp);MPI_Address( particle.d, disp+1);MPI_Address( particle.b, disp+2);MPI_Type_struct( 3, block, disp, type, &Particletype);MPI_Type_commit( &Particletype);MPI_Send( MPI_BOTTOM, 1, Particletype, dest, tag, comm);

User code for constructing an MPI datatype for a C struct


Objectives of our interface:

Automatic generation of MPI datatypes Type checking and error reporting to both sender and

receiver Seamless integration within a C program Facility to send user-defined datatypes such as

structures Facility to send multiple data in one send call (allowing

arbitrary grouping of data to be sent)


Design Overview

Network Communication

1.User Application2. Wrapper call code

3. Runtime Typed MPI

4. MPI Internals

Structure of an application that uses Typed MPI (layers2 and 3 represent Typed MPI)

1.User Application2. Wrapper call code

3. Runtime Typed MPI

4. MPI Internals

Code generated at compile time. It will perform: • signature generation• data and signature packaging• signature checking (receiver end)• error reporting

Support for runtime

Format of calls for blocking Send / Receive: int MPI_Wrapper_Send (int dest, int tag, MPI_Comm comm, (void*) buf1, int count1,…);int MPI_Wrapper_Recv (int source, int tag, MPI_Comm comm, MPI_Status *status, (void*)buf1, int count1, ...);


How does Typed MPI work ?

Phase 1: preprocess the user files determine the types of the wrapper

call parameters generate signature generate code for packaging data generate code for

sending/receiving data generate code for error

checking/reporting

Phase 2: runtime Execute the program, make the

actual MPI calls and do error checking and reporting

Sender

Construct signaturePack dataSend data(MPI_Send)Receive ack. fromreceiver

Receiver

Construct signature

Receive data(MPI_Recv)Unpack data

Compare receivedsignature with generatedoneSend data/report error toreceiver

Logical sequence of steps performed for completion of aMPI_Wrapper_Send-MPI_Wrapper_Recv operation

Time

Time


Rules for type checking. Two types T1 and T2 are identical if:• T1 and T2 have the same primitive type (e.g. int, char )• T1 and T2 are arrays and they have the same base type and the same size

(int p[5] and char q[5] are different)• T1 and T2 are structures and their members have the same type

A Signature is • generated for each variable to be sent/received based on its type• sent by the sender along with the data• compared at the receiver’s end with the expected signature (an error code will be returned to the sender )

Send / Receive Protocol

Type Checking

Receive Success/Error

SenderReceiver

Send data

Code

Receive Data

Check Type

Return Error / SuccessCode


Signatureencompasses entire type of a variable

struct example1 {int a;float b[5]; }

VAR_START

STRUCT_START

TYPE_INT

TYPE_INT TYPE_ARRAY

TYPE_FLOAT

1

1

STRUCT_START

5

struct eg1 { int a; struct eg2 { int b; float b[5] }}

• depth first approach to signature construction

• type of a variable is broken down recursively to generate the signature

{ START, VAR_START, STRUCT_START,

TYPE_INT, 1 , TYPE_ARRAY, TYPE_FLOAT, 5,

ST_UN_END, END }

Recongnition of a complex C struct

Signature


Preprocessing steps

/ * Initial user code */....MPI_Wrapper_Send (dest, tag, comm, &a, 1, &b, 2);....MPI_Wrapper_Recv (src, tag, comm, &status, &a, 1, &b, 2);...

/* Modified user source code */....MPI_Wrapper_Send79205397130821 (dest, tag, comm, &a, 1, &b, 2);....MPI_Wrapper_Recv79205397345131(src, tag, comm, &status, &a, 1, &b, 2);...

/* generated file 1 */int MPI_Wrapper_Send79205397130821 (int, int, MPI_Comm, void *, int, void * int) {/* signature and data packaging, error reporting */ }

/* generated file 2 */int MPI_Wrapper_Recv792053971345131 (int, int, MPI_Comm, MPI_Status *, void *, int, void *, int) { /* signature and data packaging, error reporting */ }

Preprocessor and generator

Compilation

Files resulting after preprocessing and source code generation

Executable

A new file is generated for every MPI_Wrapper call and it contains code for that call


Performance Evaluation

Ping Pong for blocking communication

1

10

100

1000

10000

50 100 200 400 800 1600 3200

Data size

Tim

e in

mic

rose

c

Ordinary MPI

Wrapper Calls

Ping Pong for non-blocking communication

1

10

100

1000

10000

100 200 400 800 1600 3200

Data size

Tim

e in

mic

ros

ec

Ordinary MPI

Wrapper Calls

Operation Time (inmillisec)

Preprocessing (without call to generate the code) 120

Generation of code for Send / Receive of a primitive data 80

Generation of code for Send / Receive of a structure 90

Type Checking overhead

Preprocessing overhead •The graphs and figure show that our design is feasible for automatic type generation and type checking.

•The overhead of type checking gets overshadowed by data transmission time as data size increases.


Conclusions

• A simple interface for sending user defined datatypes and multiple data in one call

• Strict type checking and error reporting to both sender and receiver

• Low preprocessing time and runtime overhead

• Error reporting to sender can be disabled to (if required) to reduce number of data transfers

Future Work

• Extensions for Group Communication and Vector operations

• Handling of expressions in MPI_Wrapper calls

• Feasibility study for use of hashing functions to reduce size of signature

Documents

Typed MPI - A Data Type Tool for MPI