CS 179: GPU ProgrammingLecture 14: Inter-process Communication

The Problem What if we want to use GPUs across a distributed system?

GPU cluster, CSIRO

Distributed System A collection of computers

Each computer has its own local memory! Communication over a network

Communication suddenly becomes harder! (and slower!)

GPUs can’t be trivially used between computers

We need some way to communicate between processes…

Message Passing Interface (MPI) A standard for message-passing Multiple implementations exist Standard functions that allow easy communication of data

between processes

Works on non-networked systems too! (Equivalent to memory copying on local system)

MPI Functions There are seven basic functions:

MPI_Init initialize MPI environment MPI_Finalize terminate MPI environment

MPI_Comm_size how many processes we have running MPI_Comm_rank the ID of our process

MPI_Isend send data (nonblocking) MPI_Irecv receive data (nonblocking)

MPI_Wait wait for request to complete

MPI Functions Some additional functions:

MPI_Barrier wait for all processes to reach a certain point

MPI_Bcast send data to all other processes MPI_Reduce receive data from all processes and

reduce to a value

MPI_Send send data (blocking) MPI_Recv receive data (blocking)

The workings of MPI Big idea:

We have some process ID We know how many processes there are

A send request/receive request pair are necessary to communicate data!

MPI Functions – Init int MPI_Init( int *argc, char ***argv )

Takes in pointers to command-line parameters (more on this later)

MPI Functions – Size, Rank int MPI_Comm_size( MPI_Comm comm, int *size ) int MPI_Comm_rank( MPI_Comm comm, int *rank )

Take in a “communicator” For us, this will be MPI_COMM_WORLD

– essentially, our communication group contains all processes Take in a pointer to where we will store our size (or rank)

MPI Functions – Send int MPI_Isend(const void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm, MPI_Request *request)

Non-blocking send Program can proceed without waiting for return! (Can use MPI_Send for “blocking” send)

MPI Functions – Send int MPI_Isend(const void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm, MPI_Request *request)

Takes in… A pointer to what we send Size of data Type of data (special MPI designations MPI_INT, MPI_FLOAT, etc) Destination process ID

…

MPI Functions – Send int MPI_Isend(const void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm, MPI_Request *request)

Takes in… (continued) A “tag” – nonnegative integers 0-32767+

Uniquely identifies our operation – used to identify messages when many are sent simultaneously

MPI_ANY_TAG allows all possible incoming messages

A “communicator” (MPI_COMM_WORLD) A “request” object

Holds information about our operation

MPI Functions - Recv int MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Request *request)

Non-blocking receive (similar idea) (Can use MPI_Irecv for “blocking” receive)

MPI Functions - Recv int MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Request *request)

Takes in… A buffer (pointer to where received data is placed) Size of incoming data Type of data (special MPI designations MPI_INT, MPI_FLOAT, etc) Sender’s process ID

…

MPI Functions - Recv int MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int source, int tag, MPI_Comm comm, MPI_Request *request)

Takes in… A “tag”

Should be the same tag as the tag used to send message! (assuming MPI_ANY_TAG not in use)

A “communicator” (MPI_COMM_WORLD) A “request” object

Blocking vs. Non-blocking MPI_Isend and MPI_Irecv are non-blocking communications

Calling these functions returns immediately Operation may not be finished!

Should use MPI_Wait to make sure operations are completed Special “request” objects for tracking status

MPI_Send and MPI_Recv are blocking communications Functions don’t return until operation is complete Can cause deadlock!

(we won’t focus on these)

MPI Functions - Wait int MPI_Wait(MPI_Request *request, MPI_Status *status)

Takes in… A “request” object corresponding to a previous operation

Indicates what we’re waiting on A “status” object

Basically, information about incoming data

MPI Functions - Reduce int MPI_Reduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)

Takes in… A “send buffer” (data obtained from every process) A “receive buffer” (where our final result will be placed) Number of elements in send buffer

Can reduce element-wise array -> array Type of data (MPI label, as before)

…

MPI Functions - Reduce int MPI_Reduce(const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)

Takes in… (continued) Reducing operation (special MPI labels, e.g. MPI_SUM, MPI_MIN) ID of process that obtains result MPI communication object (as before)

MPI Example

int main(int argc, char **argv) { int rank, numprocs; MPI_Status status; MPI_Request request; MPI_Init(&argc,&argv); MPI_Comm_size(MPI_COMM_WORLD,&numprocs); MPI_Comm_rank(MPI_COMM_WORLD,&rank); int tag=1234; int source=0; int destination=1; int count=1; int send_buffer; int recv_buffer; if(rank == source){ send_buffer=5678; MPI_Isend(&send_buffer,count,MPI_INT,destination,tag, MPI_COMM_WORLD,&request); } if(rank == destination){ MPI_Irecv(&recv_buffer,count,MPI_INT,source,tag, MPI_COMM_WORLD,&request); } MPI_Wait(&request,&status); if(rank == source){ printf("processor %d sent %d\n",rank,recv_buffer); } if(rank == destination){ printf("processor %d got %d\n",rank,recv_buffer); } MPI_Finalize(); return 0;}

Two processes

Sends a number from process 0 to process 1

Note: Both processes are running this code!

Solving Problems Goal: To use GPUs across multiple computers!

Two levels of division: Divide work between processes (using MPI) Divide work among a GPU (using CUDA, as usual)

Important note: Single-program, multiple-data (SPMD) for our purposes

i.e. running multiple copies of the same program!

Multiple Computers, Multiple GPUs Can combine MPI, CUDA streams/device selection,

and single-GPU CUDA!

MPI/CUDA Example Suppose we have our “polynomial problem” again

Suppose we have n processes for n computers Each with one GPU

(Suppose array of data is obtained through process 0)

Every process will run the following:

MPI/CUDA ExampleMPI_Init

declare rank, numberOfProcessesset rank with MPI_Comm_rankset numberOfProcesses with MPI_Comm_size

if process rank is 0:array <- (Obtain our raw data)for all processes idx = 0 through numberOfProcesses - 1:Send fragment idx of the array to process ID idx using MPI_IsendThis fragment will have the size(size of array) / numberOfProcesses)

declare local_dataRead into local_data for our process using MPI_Irecv

MPI_Wait for our process to finish receive request

local_sum <- ... Process local_data with CUDA and obtain a sum

declare global_sumset global_sum on process 0 with MPI_reduce(use global_sum somehow)

MPI_Finalize

MPI/CUDA – Wave Equation Recall our calculation:

𝑦 𝑥 ,𝑡+1=2 𝑦 𝑥, 𝑡− 𝑦𝑥 ,𝑡 −1+(𝑐∆ 𝑡∆ 𝑥 )2

(𝑦¿¿ 𝑥+1 ,𝑡−2 𝑦𝑥 ,𝑡+𝑦𝑥−1 ,𝑡 )¿

MPI/CUDA – Wave Equation Big idea: Divide our data array between n processes!

MPI/CUDA – Wave Equation In reality: We have three regions of data at a time

(old, current, new)

MPI/CUDA – Wave Equation Calculation for timestep t+1 uses the following data:

𝑦 𝑥 ,𝑡+1=2 𝑦 𝑥, 𝑡− 𝑦𝑥 ,𝑡 −1+(𝑐∆ 𝑡∆ 𝑥 )2

(𝑦¿¿ 𝑥+1 ,𝑡−2 𝑦𝑥 ,𝑡+𝑦𝑥−1 ,𝑡 )¿

tt-1

t+1

MPI/CUDA – Wave Equation Problem if we’re at the boundary of a process!

𝑦 𝑥 ,𝑡+1=2 𝑦 𝑥, 𝑡− 𝑦𝑥 ,𝑡 −1+(𝑐∆ 𝑡∆ 𝑥 )2

(𝑦¿¿ 𝑥+1 ,𝑡−2 𝑦𝑥 ,𝑡+𝑦𝑥−1 ,𝑡 )¿

x

Where do we get ? (It’s outside our process!)

tt-1

t+1

Wave Equation – Simple Solution After every time-step, each process gives its leftmost and

rightmost piece of “current” data to neighbor processes!

Proc0 Proc1 Proc2 Proc3 Proc4

Wave Equation – Simple Solution

Pieces of data to communicate:

Proc0 Proc1 Proc2 Proc3 Proc4

Wave Equation – Simple Solution Can do this with MPI_Irecv, MPI_Isend, MPI_Wait:

Suppose process has rank r: If we’re not the rightmost process:

Send data to process r+1 Receive data from process r+1

If we’re not the leftmost process: Send data to process r-1 Receive data from process r-1

Wait on requests

Wave Equation – Simple Solution Boundary conditions:

Use MPI_Comm_rank and MPI_Comm_size Rank 0 process will set leftmost condition Rank (size-1) process will set rightmost condition

Simple Solution – Problems Communication can be expensive!

Expensive to communicate every timestep to send 1 value!

Better solution: Send some m values every m timesteps!