Cilk Presentation CSCI585 Cluster Programming

1
Cilk PresentationCSCI585 Cluster Programming

• Presented by
• Asser Essam

2
Outline

• Overview
• Cilk Programming Language
• Program structure
• Memory usage
• Stack Heap Memory
• Shared Memory
• Cilk Features
• Cilk Runtime system
• Multithreaded Computational Model
• Work Stealing
• Cilk Scheduling technique

3
Overview

• A programming Language for multithreaded parallel
  programming based on ANSI C
• Designed for Shared Memory multi Processors
  (SMPs)
• It is algorithmic in that the runtime systems
  scheduler guaranties efficient and predictable
  performance
• It allows the programmer to concentrate on the
  program logic and its runtime system takes care
  of the other details as load balancing and
  communication protocols
• It doesnt use message passing techniques

4
Cilk Programming Language Program Structure

• Procedure is the building block of a Cilk
  program.
• It is like an ordinary C function with new
  keywords.
• The Cilk keywords are only used within a Cilk
  procedure.

include ltstdli.hgt include ltstdio.hgt include
ltcilk.hgt cilk int fib (int n) if (nlt2) return
n else int x,y x spawn fib
(n-1) y spawn fib(n-2)
sync return (xy)
5
Cilk Programming LanguageProgram Structure
Cont..

• A Cilk program execution can be viewed as a
  directed acyclic graph.
• A procedure is considered as a sequence of non
  blocking threads represented horizontally in the
  graph
• Each thread can spawn a new thread as represented
  by the vertical lines
• Each thread must wait for its children to return
  their results in order for it to return.

6
Cilk Programming Language Program Structure
Cont..

• The command cilk -02 fib.cilk o fib compiles
  the program and outputs the object file ready for
  execution
• The command fib nproc 4 30 runs the program on
  4 processes giving it the argument 30
• Using different flags when you run the program
  can generate important statistics like --stats
  which generates the time taken for execution by
  each processor.

7
Cilk Programming Language Stack Heap Memory

• As in C Cilk has two types of memory, Heap
  memory and stack memory
• In Cilk the stack is called cactus stack and this
  is because more than one procedure access it
  (parallel)
• Each procedure sees it a different view of the
  stack
• It has the same limitations as the stack in C
  i.e. a sibling cant pass a variable that it has
  allocated to its parent

8
Cilk Programming LanguageShared Memory

• Cilk supports shared memory
• Shared memory can occur when parallel procedures
  access global objects or indirectly when we pass
  a pointer to a spawned procedure
• According to the logic of the code, the use of
  shared memory may lead to race conditions
• See examples on next slide

9
Cilk Programming LanguageShared Memory Cont..
cilk int foo (void) int x 0, y spawn
bar(x) x x 1 sync return
(x) cilk void bar (int px)
px px 1 return
cilk int foo (void) int x 0, y spawn
bar(x) y x 1 sync return
(y) cilk void bar (int px)
printf("d", px 1) return
10
Cilk Programming Language Cilk Features

• Locking Cilk provides mutual exclusive locks
  through the use of critical sections of the code
• Inlets This feature allows you to use the value
  returned from a child thread in an expression
• Aborting A procedure can cancel unnecessary
  spawned work
• Interacting with Cilk scheduler This is
  important to make use of already allocated memory
  instead of using new one (using SYNCHED)

11
Cilk Runtime systemMultithreaded Computational
Model

• A Cilk procedure can be viewed as a dag
• This means that there are dependencies between
  the different threads
• These dependencies form a partial order
  permitting many ways to schedule the threads in
  the dag
• The efficiency of the computation depends on the
  way that the dag is mapped to the underlying
  processors
• Each Cilk program produces a well structured dag
  that can be scheduled efficiently

12
Cilk Runtime systemWork-Stealing

• The Cilk runtime system implements a scheduling
  policy that depends on work-stealing
• When a processor runs out of work it randomly
  picks another processor and takes work from it
• To understand how this works, lets remember what
  happens in the computation of a C program
• When a call to a child is done the state of the
  parent is pushed on the stack
• Then when the call to the child is processed, the
  state of the parent is popped and computation
  continues at the parent level

13
Cilk Runtime systemWork-Stealing Cont..

• Work stealing is done using the following
  algorithm
• Since a Cilk procedure performs serially on one
  processor as a c program
• Locally, a processor executes procedures in a
  serial manner exploring the spawn tree in a depth
  first manner
• When a child is spawned the state of the parent
  is pushed to the stack and the processor starts
  working on the child
• When another processor request work, it steals
  work from the other end of the stack which in
  this case will be far from the work area of the
  victim processor

14
Cilk Runtime systemCilk Scheduling technique

• The Cilk language has two types of schedulers
• Nanoscheduler its job is to schedule procedures
  within a single processor
• Microscheduler its job is to schedule procedures
  on multiple processors

15
Cilk Runtime systemNanoscheduler

• The nanoscheduler has to make decisions very
  quickly and this is why it is compiled within the
  code by the Cilk2c compiler
• It is very easy to implement as it converts each
  spawn call to its equivalent procedure call and
  each synch call to no operation
• In order to be used with the microscheduler we
  need to keep track of the current scheduling
  state
• The nanoscheduler uses a deque of frames for this
  purpose

16
Cilk Runtime systemNanoscheduler Cont..

• A cilk frame is a data structure that can hold
  the state of the procedure
• A deque is a double ended queue
• A deque is considered as a stack from the
  nanoscheduler perspective
• The deque is tightly coupled to the C stack with
  a one to one correspondence between the frames in
  the deque and the activation frames on the c
  stack
• When a procedure is first called, it initializes
  a frame and pushes it on the stack
• When a spawn occurs, the state of the current
  procedure is saved on the stack
• When the procedure completes, its frame is popped

17
Cilk Runtime systemMicroscheduler

• The microschedulers job is to schedule
  procedures across a fixed set of processors
• It is implemented as a randomized work stealing
  scheduler
• When a processor runs out of work, it picks a
  processor randomly and checks if its deque has
  frames to be computed
• It takes the oldest frame and places the
  corresponding procedure on its nanscheduler

18
Cilk Runtime systemMicroscheduler Cont..

• The stolen procedure is called the slow version
  of the procedure
• As being invoked from a generic scheduling
  routine, it has a standard interface
• It has lots of overheads as
• It has a pointer to a frame as its input variable
  instead of having the data directly. It has to
  extract the input data
• It has to pack the results instead of returning
  them directly