concurrency scalability

Mårten RångeWCOM AB

@marten_range

ConcurrencyExamples for .NET

Responsive

PerformanceScalable algorithms

Three pillars of Concurrency

Scalability (CPU) Parallel.For

Responsiveness Task/Future async/await

Consistency lock/synchronized Interlocked.* Mutex/Event/Semaphore Monitor

Scalability

Which is fastest?

var ints = new int[InnerLoop];var random = new Random();for (var inner = 0; inner < InnerLoop; ++inner){    ints[inner] = random.Next();}// ------------------------------------------------var ints = new int[InnerLoop];var random = new Random();Parallel.For(    0,     InnerLoop,    i => ints[i] = random.Next()    );

SHARED STATE Race condition

var ints = new int[InnerLoop];var random = new Random();for (var inner = 0; inner < InnerLoop; ++inner){    ints[inner] = random.Next();}// ------------------------------------------------var ints = new int[InnerLoop];var random = new Random();Parallel.For(    0,     InnerLoop,    i => ints[i] = random.Next()    );

SHARED STATE Poor performancevar ints = new int[InnerLoop];var random = new Random();for (var inner = 0; inner < InnerLoop; ++inner){    ints[inner] = random.Next();}// ------------------------------------------------var ints = new int[InnerLoop];var random = new Random();Parallel.For(    0,     InnerLoop,    i => ints[i] = random.Next()    );

Then and now

Metric VAX-11/750 (’80)

Today Improvement

MHz 6 3300 550x

Memory MB 2 16384 8192x

Memory MB/s 13 R ~10000W ~2500

770x190x

Then and now

Metric VAX-11/750 (’80)

Today Improvement

MHz 6 3300 550x

Memory MB 2 16384 8192x

Memory MB/s 13 R ~10000W ~2500

770x190x

Memory nsec 225 70 3x

Then and now

Metric VAX-11/750 (’80)

Today Improvement

MHz 6 3300 550x

Memory MB 2 16384 8192x

Memory MB/s 13 R ~10000W ~2500

770x190x

Memory nsec 225 70 3x

Memory cycles

1.4 210 -150x

299,792,458 m/s

Speed of light is too slow

0.09 m/c

99% - latency mitigation

1% - computation

2 Core CPU

RAM

L3L2

L1

CPU1

L2

L1

CPU2

2 Core CPU – L1 Cache

L1

CPU1

L1

CPU2

new Random ()

new int[InnerLoop]

2 Core CPU – L1 Cache

L1

CPU1

L1

CPU2

Random object Random object

2 Core CPU – L1 Cache

L1

CPU1

L1

CPU2

Random object Random object

2 Core CPU – L1 Cache

L1

CPU1

L1

CPU2

Random objectRandom object

2 Core CPU – L1 Cache

L1

CPU1

L1

CPU2

Random objectRandom object

2 Core CPU – L1 Cache

L1

CPU1

L1

CPU2

Random objectRandom object

2 Core CPU – L1 Cache

L1

CPU1

L1

CPU2

Random objectRandom object

4 Core CPU – L1 Cache

L1

CPU1

L1

CPU2

L1

CPU3

L1

CPU4

new Random ()

new int[InnerLoop]

2x4 Core CPU

RAM

L3L2

L1

CPU1

L2

L1

CPU2

L2

L1

CPU3

L2

L1

CPU4

L3L2

L1

CPU5

L2

L1

CPU6

L2

L1

CPU7

L2

L1

CPU8

Solution 1 – Locks

var ints = new int[InnerLoop];var random = new Random();Parallel.For(    0,     InnerLoop,    i => {lock (ints) {ints[i] = random.Next();}}    );

Solution 2 – No sharing

var ints = new int[InnerLoop];Parallel.For( 0, InnerLoop, () => new Random(), (i, pls, random) => {ints[i] = random.Next(); return random;}, random => {} );

Parallel.For adds overheadLevel0

Level1

Level2

ints[0]

ints[1]

Level2

ints[2]

ints[3]

Level1

Level2

ints[4]

ints[5]

Level2

ints[6]

ints[7]

Solution 3 – Less overhead

var ints = new int[InnerLoop];Parallel.For( 0, InnerLoop / Modulus, () => new Random(), (i, pls, random) => { var begin = i * Modulus ; var end = begin + Modulus ; for (var iter = begin; iter < end; ++iter) { ints[iter] = random.Next(); } return random; }, random => {} );

var ints = new int[InnerLoop];var random = new Random();for (var inner = 0; inner < InnerLoop; ++inner){    ints[inner] = random.Next();}

Solution 4 – Independent runs

var tasks = Enumerable.Range (0, 8).Select ( i => Task.Factory.StartNew ( () => { var ints = new int[InnerLoop]; var random = new Random (); while (counter.CountDown ()) { for (var inner = 0; inner < InnerLoop; ++inner) { ints[inner] = random.Next(); } } }, TaskCreationOptions.LongRunning)) .ToArray ();Task.WaitAll (tasks);

Parallel.For

Only for CPU bound problems

Sharing is bad

Kills performanceRace conditions

Dead-locks

Cache locality

RAM is a misnomerClass designAvoid GC

Natural concurrency

Avoid Parallel.For

Act like an engineer

Measure before and after

One more thing…

http://tinyurl.com/wcom-cpuscalability

Mårten RångeWCOM AB

@marten_range