Multi-Tasking Map (MapReduce, Tasks in Rust)

cs4414 Fall 2013
University of Virginia
David EvansJodhpur, India (Dec 2011)

Plan for Today
• Recap list map
• Google’s MapReduce
• Tasks in Rust
• Multi-threaded map
26 September 2013 University of Virginia cs4414 1
PS2 is due Monday (30 Sept) at 8:59pm.
Submission form will be posted later today, and
include signup for scheduling your demo/review.
All team members are expected to participate in
the review, except in extreme circumstances.

struct Node {
head : int,
tail : Option<~Node>
}
type List = Option<~Node> ;
trait Map {
fn mapr(&self, &fn(int) -> int) -> List;
}
impl Map for List {
fn mapr(&self, f: &fn(int) -> int) -> List {
match(*self) {
None => None,
Some(ref node) => { Some(~Node{ head: f(node.head),
tail: node.tail.mapr(f) }) },
} } }
You should understand everything in this code.
Ask questions now if there is anything unclear.

Cost of Map
Core 1
What is the running time of p.map(f)
using one core where p is a list of N
elements and each evaluation of f(x)
takes 1ms?

Cost of Multi-Core Map
Core 1
Core 3
Core 2
Core 4
What is the running time of p.map(f)
using k cores where p is a list of N
elements and each evaluation of f(x)
takes 1ms?

How should we parallelize map?
match(*self) {
None => None,
Some(ref node) => { Some(~Node{ head: f(node.head),
tail: node.tail.mapr(f) }) },
}
}

“MapReduce is a programming model and an associated
implementation for processing and generating large data sets. Users
specify a map function that processes a key/value pair to generate a
set of intermediate key/value pairs, and a reduce function that
merges all intermediate values associated with the same
intermediate key. Many real world tasks are expressible in this
model, as shown in the paper.
Programs written in this functional style are automatically
parallelized and executed on a large cluster of commodity machines.”
OSDI 2004

Did
Google
invent
map?

10
1955-1960: First “mass-produced” computer (sold 123 of them)
1 accumulator register (38 bits), 3 decrement registers (15 bit)
Instructions had 3 bit opcode, 15 bit decrement, 15 bit address
Magentic Core Memory
32,000 36-bit words
40,000 instructions/second

11
John McCarthy
playing chess
with IBM 7090
(1967)

match(*self) {
None => None,
Some(ref node) => {
Some(~Node{ head: f(node.head), tail: node.tail.mapr(f) }) },
}
}

@ pointers (in 1960)

MapReduce
Google’s map:
match(*self) {
None => None,
Some(ref node) => {
Some(~Node{ head: f(node.head),
tail: node.tail.mapr(f) }) }, } }

fn mapg<K1, V1, K2, V2>(List<Pair<K1, V1>>,
f: &fn(K1, V1) -> (K2, V2))
-> List<Pair<K2, V2>>
fn reduceg<K, V, R>(K, List<V>) -> List<R>

fn mapg<K1, V1, K2, V2>(List<Pair<K1, V1>>,
f: &fn(K1, V1) -> (K2, V2))
-> List<Pair<K2, V2>>
fn reduceg<K, V, R>(K, List<V>) -> List<R>
fn map_reduce<K1, V1, K2, V2, R>(
List<Pair<K1, V2>>,
mapf: &fn(K1, V1) -> (K2, V2)),
reducef: &fn(K2, List<V2>) -> R))
-> List<R>

data: List<Pair<K1, V2>>,
mapf: &fn(K1, V1) -> (K2, V2)),
-> List<R> {
}

data: List<Pair<K1, V2>>,
mapf: &fn(K1, V1) -> (K2, V2)),
-> List<R> {
let ivalues = data.map(mapf)
let mvalues = // merge ivalues by k2
mvalues.map(reducef)
}
Completing the code (with parallel map will finish
today) is left as sticker-worthy exercise!

Mapping in
Parallel

Processes, Threads, Tasks
Process
Originally: abstraction
for owning the whole
machine
What do you need:
Thread
(Illusion of)
independent sequence
of instructions
What do you need:

Processes, Threads, Tasks
Process
Originally: abstraction
for owning the whole
machine
What do you need:
Own program counter
Own stack, registers
Own memory space
Own program counter
Shares memory space
Thread
(Illusion of)
independent sequence
of instructions
What do you need:

Tasks in Rust

Tasks
Own PC
Safely shared
immutable memory
Safely independent
own memory
fn spawn(f: ~fn())
spawn( | | {
println(“Get back to work!”);
});
do spawn {
println(“Get back to work!”);
}
syntactic sugar:
Task = Thread – unsafe memory sharing
or
Task = Process + safe memory sharing – cost of OS process

impl Map for List {
match(*self) {
None => None,
Some(ref node) => {
Some(~Node{ head: f(node.head),
tail: node.tail.mapr(f) })
},
}
}
}
Original single-threaded mapr
fn spawn(f: ~fn())

impl Map for List {
fn mapr(&self, f: extern fn(int) -> int) -> List {
match(*self) {
None => None,
Some(ref node) => {
do spawn {
f(node.head)
}
Some(~Node{ head: ?,
},
}
}
}
First attempt
Cannot use node here!

impl Map for List {
match(*self) {
None => None,
Some(ref node) => {
let val = node.head;
do spawn {
f(val)
}
Some(~Node{ head: ?,
},
}
}
}
How can we get results back from a spawned task without shared memory?

Channels
let (port, chan) : (Port<int>, Chan<int>) = stream();
do spawn {
chan.send(f(val));
}
let newval = port.recv();

Using streams to
spawn is dangerous
for salmon, but Rust
saves you from (data)
races with the bears!

First attempt
match(*self) {
None => None,
Some(ref node) => {
let newtail = node.tail.mapr(f);
do spawn {
chan.send(f(val));
}
Some(~Node{ head: port.recv(), tail: newtail })
}
}
}
} Compiles are runs fine and produces correct output…
but has a major bug!

Now we’re spawning!
match(*self) {
None => None,
Some(ref node) => {
do spawn {
chan.send(f(val));
}
let newtail = node.tail.mapr(f);
Some(~Node{ head: port.recv(), tail: newtail })
}
}
}
}

fn collatz_steps(n: int) -> int {
if n == 1 { 0 } else { 1 + collatz_steps(if n % 2 == 0 { n / 2 } else { 3*n + 1 }) }
}
fn find_collatz(k: int) -> int {
// Returns the minimum value, n, with Collatz stopping time >= k.
let mut n = 1;
while collatz_steps(n) < k { n += 1; }
n
}
fn main() {
let lst0 : List = Some(~Node{head: 400, tail: .
Some(~Node{head : 410, tail:
// … 16 total similar elements
} );
println(lst0.to_str());
let lst1 = lst0.mapr(find_collatz);
let lst2 = lst1.mapr(find_collatz);
}

When 350+% of your CPU isn’t fast enough, its time to buy a new computer!

Intel i7 Quad-Core Processor

Intel i7 Quad-Core Processor
Core Core Core Core
Shared Memory Cache (L3 = 6MB)
~256KBL2
Cache(?)

Why so few?

Portuguese, was beavering away in the library when ‘smoke
suddenly started to come out’ of her computer. Fortunately, she
removed the fire hazard from the library, averting disaster at the
last moment. The student gave The Tab her version of the story:
“I was in the library working at my computer when smoke
suddenly started to come out of it. I freaked out for a second,
trying to save my work onto my hard disk, but then I realised it
was probably more important to take it out of the library.
The Tab (Oxford), “Laptop Fire Almost Destroys College Library”

Where the Cores Are
nVIDIA GeForce GTX 650M
384 cores
(but even harder for
typical programs to
use well than Intel’s
cores)

How much faster will my Rust mapping
program be on my new machine?
2013 MacBook Pro
Intel i7-3740QM
2.7 GHz, 4 cores (8 threads)
6MB shared L3 cache
2011 MacBook Air
Intel i5-2557M
1.7 GHz, 2 cores (4 threads)
3 MB shared L3 cache
both support “hyperthreading” (two threads per core)
60 seconds
(normalized time,
running on 16-
element list)
?

Submit your “guesses” and reasoning
in course forum….hopefully I will
know the actual answer by Tuesday!
PS2 is due Monday (30 Sept) at 8:59pm.
Submission form will be posted later today, and
include signup for scheduling your demo/review.
All team members are expected to participate in
the review, except in extreme circumstances.

Multi-Tasking Map (MapReduce, Tasks in Rust)

More Related Content

What's hot

Viewers also liked

Similar to Multi-Tasking Map (MapReduce, Tasks in Rust)

More from David Evans

Recently uploaded

Multi-Tasking Map (MapReduce, Tasks in Rust)