It Aint the Meat its the Notion Why

It Ain’t the Meat, it’s the Notion Why Theory is Essential to Teaching Concurrent Programming Maurice Herlihy Brown University Multicore Programming Education

In many areas of CS • Systems research fought all the hard battles – FORTRAN complexity theory – UDP/IP network algorithms • Theory researchers arrived later to shoot the wounded • But concurrency is different Multicore Programming Education 2

Vive la Différence • Concurrent programmers are made, not born (mostly) • Students have little experience with concurrency (today) • Develop intuition by – Idealized examples first – Logical reasoning Multicore Programming Education 3

Homework Challenge Devise Wait-Free 1 Enqueuer, 1 -Dequeuer Buffer using only reads & writes Multicore Programming Education 4

Another Homework Challenge How about two dequeuers? Multicore Programming Education 5

Shared Memory Atomic reads and writes to individual variables write read Multicore Programming Education 6

Snapshot Challenge Wait-Free algorithm to read multiple variables atomically? Read both Multicore Programming Education 7

Snapshot Challenge Wait-Free algorithm to write multiple variables atomically? write both Multicore Programming Education 8

Bakery Algorithm class Bakery implements Lock { boolean[] flag; 6 Label[] label; 2 0 n-1 public Bakery (int n) { flag = new boolean[n]; f f t f f label = new Label[n]; 0 0 4 0 0 5 0 0 for (intfirst-comei = 0; i < n; i++) { Classic, false; label[i] = 0; firstflag[i] served= mutual } exclusion algorithm } … Critical Section Multicore Programming Education 9

Is Locked? is. Locked() True means was locked at some instant, false means was free at some instant. Multicore Programming Education Bakery Lock 10

I Can Has Lock? try. Lock() Acquire lock if free & return true, otherwise immediately return false. Multicore Programming Education Bakery Lock 11

Our Experience Teaching Concurrency • Brown (10 years) • Tel-Aviv (10 years) • Industry Multicore Programming Education 12

Mutual Exclusion • Peterson, Bakery, space lower bounds • Teaches – Deadlock – Livelock – Fairness, first-come-first-served, … Multicore Programming Education 13

Linearizability • How to describe a concurrent object? – By “equivalence” to sequential • What is a concurrent API? • How do we reason about correctness? – Even informally! • Learn the rules before you break them Multicore Programming Education 14

Consensus Numbers • Certain synchronization primitives are mathematically stronger than others • No, you can’t build lock-free X from Y – So do not waste your time trying • Yes, you can build lock-free X from Z – The rest is optimization Multicore Programming Education 15

Uh, Oh, Spin Locks bad Test&test&set less bad Test&set • Inexplicable in idealized model • Well, kids … – We didn’t actually lie to you – But the architects gave us caches – And you have to outwit them … Multicore Programming Education 16

Then What? • Implement simple T&S lock – Dreadful performance • Implement T&T&S lock – Improved performance Multicore Programming Education 17

Homework Challenge Devise Wait-Free 1 Enqueuer, 1 -Dequeuer Buffer using only reads & writes Multicore Programming Education 18

Homework Challenge , s e Y u o y n a c Multicore Programming Education 19

Another Homework Challenge How about two dequeuers? Multicore Programming Education 20

Maybe? “hence [any wait-free object] can be written without a CAS/LL/SC. Instead LL/SC is used to make it more efficient but LL/SC isn't *required* to make it wait-free. ” Found in a blog Multicore Programming Education 21

Another Homework Challenge Al as , n o, yo uc an Multicore Programming Education ’t 22

What If we have Test&Set? Multicore Programming Education 23

What If we have Test&Set? No , y ou s til l ca n’t Multicore Programming Education 24

What If we have Compare&Swap? Multicore Programming Education 25

What if We Have Compare&Sswap? , s e Y u o y n a c Multicore Programming Education 26

Snapshot Challenge Wait-Free algorithm to read multiple variables atomically? Reads from 0 and 1 Reads from 1 and 2 Multicore Programming Education 27

Snapshot Challenge Wait-Free algorithm to read multiple variables atomically? , s e Y u o y n Reads from a c 0 and 1 Reads from 1 and 2 Multicore Programming Education 28

Multiple Assignment Challenge Wait-Free algorithm to write multiple variables atomically? Writes to 0 and 1 Writes to 1 and 2 Multicore Programming Education 29

Multiple Assignment Challenge Wait-Free algorithm to write multiple No variables , atomically? yo uc Writes to 0 and 1 an ’t Writes to 1 and 2 Multicore Programming Education 30

Is Locked? is. Locked() True means was locked at some instant, false means was free at some instant. Multicore Programming Education Bakery Lock 31

Is Locked? is. Locked() u o y n a c , True means was locked at s e some instant, Y false means was free at some instant. Multicore Programming Education Bakery Lock 32

I Can Has Lock? try. Lock() Acquire lock if free & return true, otherwise immediately return false. Multicore Programming Education Bakery Lock 33

I Can Has Lock? try. Lock() No , y ou c Acquire lock if free &an ’ t return true, otherwise immediately return false. Multicore Programming Education Bakery Lock 34

My Very Last Slide • Teaching theory first – Develops intuition – Imposes order on chaos • Idealized problems – Prepare for realistic, complicated ones • Impossibility results – Save time and effort Multicore Programming Education 35