1

HOW THE CLOUD WORKS

Ken BirmanCornell University

2

Consider Facebook

Popular social networking site (currently blocked in China), with > 1B users, growing steadily

Main user interface: very visual, with many photos, icons, ways to comment on things, “like/dislike”

Rapid streams of updates

3

Facebook Page

Page itself was downloaded from Facebook.com They operate many

data centers, all can serve any user

Page was full of URLs Each URL triggered

a further download. Many fetched photos or videos

User’s wall: a continuously scrolling information feed with data from her friends,

news, etc

4

Facebook image fetching architecture

local cache

Akamai

Facebook Edge

Edge Cache

Facebook Edge

Edge

Cache

Facebook Edge

Edge

Cache

Facebook Edge

Edge

Cache

Facebook Resizer

Resizer

Cache

Haystack

Akamai

Akamai

Akamai

Cloud

5

The system...

Operates globally, on every continent Has hundreds of Facebook Edge sites Dozens of Resizer locations just in the

USA, many more elsewhere A “few” Haystack systems for each

continent

Close relationships with Akamai, other network providers

6

Things to notice

The cloud isn’t “about” one big system We see multiple systems that talk easily to

on-another all using web-page standards (XML, HTML, MPLS...)

They play complementary roles Facebook deals with Akamai for caching of

certain kinds of very popular images, AdNet and others for advertising, etc

And within the Facebook cloud are many, many, interconnected subsystems

7

Why so many caches?

To answer, need to start by understanding what a cache does A cache is a collection of web pages and

images, fetched previously and then retained

A request for an image already in the cache will be satisfied from the cache

Goal is to spread the work of filling the Facebook web page so widely that no single element could get overloaded and become a bottleneck

8

But why so many layers?

Akamai.com is a web company dedicated to caching for rapid data (mostly images) delivery If Facebook uses Akamai, why would

Facebook ever need its own caches? Do the caches “talk to each other”? Should

they?

To understand the cloud we should try and understand answers to questions like these

9

Memcached: A popular cache Stands for “In-Memory Caching Daemon”

A simple, very common caching tool Each machine maintains its own (single)

cache

function get_foo(foo_id) foo = memcached_get("foo:" . foo_id) return foo if defined foo foo =

fetch_foo_from_database(foo_id) memcached_set("foo:" . foo_id, foo) return foo

end

10

Should we use Memcached everywhere?

Cached data can be stale (old/incorrect) A cache is not automatically updated when data changes;

need to invalidate or update the entry yourself And you may have no way to know that the data changed

When a cache gets full, we must evict less popular content. What policy will be used? (Memcached: LRU)

When applications (on one machine) share Memcached, they need to agree on the naming rule they will use for content Otherwise could end up with many cached copies of Angelina

Joli and Brad Pitt, “filling up” the limited cache space

11

Fun with Memcached

There are systems built over Memcached that have become very important Berkeley Spark system is a good example Spark Memcached + a nice rule for

naming what the cache contains

Spark approach focuses on in-memory caching on behalf of the popular MapReduce/Hadoop computing tool

12

MapReduce/Hadoop

Used when searching or indexing very large data sets constructed from collections of web data For example, all the web pages in the Internet Or all the friending relationships in all of Facebook

Idea is to spread the data over many machines, then run highly parallel computations on unchanging data The actual computing tends to be simple programs Map step: spreads computing out. Reduce: combines

intermediary results. Result aggregated with exactly one copy of each intermediary output

Often iterative: second step depends on output of first step

13

Spark: Memcached for MapReduce Spark developers reasoned that if MapReduce

uses files for the intermediary results, file I/O would be a peformance limiting cost Confirmed this using experiments It also turned out that many steps recompute nearly

the identical thing (for example by counting words in a file)

Memcached can help... if MapReduce can find the precomputed results and if we “place” tasks to run where those

precomputed results are likely to be found

14

Spark “naming convention”

Key idea in Spark: rather than name intermediate results using URLs or file names, they use the “function that produced the result” Represented in a functional programming notation

based on the Microsoft LINQ syntax In effect: “This file contains ((X))”

Since the underlying data is unchanging, e.g. a file, “X” has same meaning at all times

Thus ((X)) has a fixed meaning too By cleverly keeping subresults likely to be

reused, Spark obtained huge speedups, often 1000x or more!

15

Spark has become very important While idea is simple, providing a full

range of control over what is in the cache and when it is searched and when things are evicted is complex Spark functionality was initially very limited

but has become extremely comprehensive There is a user community of tens of

thousands

In the cloud, when things take off, they “go viral”! … even software systems

16

Key insight

In-memory caching can have huge and very important performance implications

Caching, in general, is of vital importance in the cloud, where many computations run at high load and data rates and immense scale

But not every cache works equally well!

17

Back to Facebook

Seeing how important Spark became for MapReduce, we can ask questions about the Facebook caching strategy Are these caches doing a good job? What

hit rate do they achieve? Should certain caches “focus” on retaining

only certain kinds of data, while other caches specialize in other kinds of data? When a Facebook component encounters a photo or video, can we “predict” the likely value of caching a copy?

18

Using Memcached in a pool of machines

Facebook often has hundreds or thousands of machines in one spot, each can run Memcached

They asked: why not share the cache data? Leads to a distributed cache structure They built one using ideas from research

experiences

A distributed hash table offers a simple way to share data in a large collection of caches

19

How it works

We all know how a HashMap or HashTable works in a language like Java or C# or C++ You take the object you want to save and

compute a HashCode for it. This is an integer and will look “random” but is deterministic for any single object

For example, it could be the XOR of the bytes in a file

Hashcodes are designed to spread data very evenly over the range of possible integers

Network communication

It is easy for a program on biscuit.cs.cornell.edu to send a message to a program on “jam.cs.cornell.edu” Each program sets up a “network socket Each machine has an IP address, you can

look them up and programs can do that too via a simple Java utility

Pick a “port number” (this part is a bit of a hack)

Build the message (must be in binary format)

Java utils has a request

20

Distributed Hash Tables

It is easy for a program on biscuit.cs.cornell.edu to send a message to a program on “jam.cs.cornell.edu”

... so, given a key and a value1. Hash the key2. Find the server that “owns” the hashed

value 3. Store the key,value pair in a “local”

HashMap there

To get a value, ask the right server to look up key

21

22

List of machines

There are several ways to track the machines in a network. Facebook just maintains a table In each FB data center there is a big table

of all machines currently in use Every machine has a private copy of this

table, and if a machine crashes or joins, it is quickly updates (seconds)

Can we turn our table of machines into a form of HashMap?

23

From a table of machines to a DHT Take the healthy

machines Compute the

HashCode for each using its name, or ID

These are integers in range [Int.MinValue, Int.MaxValue]

Rule: an object with HashCode (O) will be be placed on the K machines closest to (O)

24

Side remark about tracking members

This Facebook approach uses a “group” of machines and a “view” of the group We will make use of this in later lectures too But it is not the only way! Many DHTs track just

log(N) of the members and build a routing scheme that takes log(N) “hops” to find an object (See: Chord, Pastry...)

The FB approach is an example of a 1-hop DHT

Cloud systems always favor 1-hop solutions if feasible

Distributed Hash Tables25

dht.Put(“ken”,2110)

(“ken”, 2110)

dht.Get(“ken”)

“ken”.hashcode()%N=77 IP.hashcode()%N=77

123.45.66.781 123.45.66.782 123.45.66.783 123.45.66.784

IP.hashcode()%N=98

IP.hashcode()%N=13

IP.hashcode()%N=175

hashmap kept by 123.45.66.782

“ken”.hashcode()%N=77

Facebook image “stack”

We decided to study the effectiveness of caching in the FB image stack, jointly with Facebook researchers

This stack’s role is to serve images (photos, videos) for FB’s hundreds of millions of active users About 80B large binary objects (“blob”) / day FB has a huge number of big and small data centers

“Point of presense” or PoP: some FB owned equipment normally near the user

Akamai: A company FB contracts with that caches images FB resizer service: caches but also resizes images Haystack: inside data centers, has the actual pictures (a

massive file system)

26

What we instrumented in the FB stack

Think of Facebook as a giant distributed HashMap Key: photo URL (id, size, hints about where

to find it...) Value: the blob itself

27

Facebook traffic for a week

Client activity varies daily....

... and different photos have very different popularity statistics

28

Facebook cache effectiveness Existing caches are very effective... ... but different layers are more effective

for images with different popularity ranks

29

Facebook cache effectiveness

Each layer should “specialize” in different content.

Photo age strongly predicts effectiveness of caching

30

Hypothetical changes to caching? We looked at the

idea of having Facebook caches collaborate at national scale…

… and also at how to

vary caching based on the “busyness” of the client

31

Social networking effect?

Hypothesis: caching will work best for photos posted by famous people with zillions of followers

Actual finding: not really

32

Locality?

Hypothesis: FB probably serves photos from close to where you are sitting

Finding: Not really...

… just the same, ifthe photo exists, itfinds it quickly

33

Can one conclude anything?

Learning what patterns of access arise, and how effective it is to cache given kinds of data at various layers, we can customize cache strategies

Each layer can look at an image and ask “should I keep a cached copy of this, or not?”

Smart decisions Facebook is more effective!

34

Strategy varies by layer

Browser should cache less popular content but not bother to cache the very popular stuff

Akamai/PoP layer should cache the most popular images, etc...

We also discovered that some layers should “cooperatively” cache even over huge distances Our study discovered that if this were done

in the resizer layer, cache hit rates could rise 35%!

35

36

… many research questions arise Can we design much better caching

solutions?

Are there periods with bursts of failures? What causes them and what can be done?

How much of the data in a typical cache gets reused? Are there items dropped from cache that should have been retained?

Overall picture in cloud computing Facebook example illustrates a style of

working Identify high-value problems that matter to

the community because of the popularity of the service, the cost of operating it, the speed achieved, etc

Ask how best to solve those problems, ideally using experiments to gain insight

Then build better solutions

37

38

Learning More?

We have a paper with more details and data in the 2013 version of ACM Symposium on Operating Systems Principles, SOSP First author is Qi Huang, a Chinese student

who created the famous PPLive system, was studying for his PhD at WUST, then came to Cornell to visit

Qi will eventually earn two PhD degrees! One awarded by Cornell, one by WUST after he finishes

A very amazing and talented cloud computing researcher

39

More about caching

Clearly, caching is central to modern cloud computing systems!

But the limitation that we are caching static data is worrying MapReduce/Hadoop use purely static data FB images and video are static data too But in “general” cloud computing will have

very dynamic kinds of data, rapidly changing

40

Coherent Cache

We say that a cache is coherent if it is always a perfect real-time replica of the “true” data True object could be in a database or file system Or we could dispense with the true object and

use only in the in-memory versions. In this case the cache isn’t really a cache but is actually an in-memory replication scheme

In the cloud, file system access is too slow! So we should learn more about coherent

caching

41

What could a coherent cache hold? A standard cache just has “objects” from

some data structure

A coherent cache could hold the entire data structure! A web graph with pages and links A graph of social network relationships, Twitter

feeds and followers, etc Objects on a Beijing street, so that a self-driving

car can safely drive to a parking area, park itself, and drive back later to pick you up

42

Coherent data replication

Clearly we will need to spread our data widely: “Partitioning” is required We partition data in space (like with a DHT) Also in time (e.g. version of the database at time T,

T+1, …) Sometimes hierarchically (e.g. users from the US North

East, US Central, US North West…) Famous paper by Jim Gray & others: Dangers of

Database Replication and a Solution Shows that good partitioning functions are critically

needed Without great partitioning, replication slows a system

down!

43

Aspects of coherent replication A partitioning method Many servers, small subsets for each partition

(“shard” has become the common term) Synchronization mechanisms for conflicting

actions A method for updating the data A way to spread the “read only” work over the

replicas Shard membership tracking Handling of faults that cause whole shard to

crash

44

Does Facebook have coherency? Experiments reveal answer: “no”

Create a Facebook account and put many images on it

Then form networks with many friends Now update your Facebook profile image a

few times

Your friends may see multiple different images of you on their “wall” for a long period of time! This reveals that it takes a long time

(hours) for old data to clear from the FB cache hierarchy!

Inconsistencies in the cloud

46

In fact this is common in today’s cloud

We studied many major cloud providing systems

Some guarantee coherency for some purposes but the majority are at best weakly coherent When data changes they need a long time

to reflect the updates They cache data heavily and don’t update

it promptly

47

CAP Theorem

Proposed by Berkeley Professor Eric Brewer “You can have just 2 of Consistency, Availability,

Partitioning or Fault Tolerance” He argues that consistency is the guarantee to

relax We will look at this more closely later

Many people adopt his CAP based views This justifies non-coherent caching But their systems can’t solve problems in

ways guaranteed to be safe

48

High assurance will need more! Remember that we are interested in the

question of how one could create a high assurance cloud!

Such a cloud needs to make promises If a car drives on a street it must not run

over people If a power system reconfigures it must not

explode the power generators A doctor who uses a hospital computing

system needs correct and current data

49

So… we must look at coherent caching

In fact we will focus on “data replication” Data that should be in memory, for speed With a few shard members holding the information But with guarantees: if you compute using it, the data is

current

Why not a full database? Our coherent replication methods would live in

structures like the Facebook infrastructure: big, complex We need to build these in ways optimized to the uses Hence databases might be elements but we can’t just

hand the whole question to a database system

50

Summary

We looked at the architecture of a typical very large cloud computing system (Facebook)

We saw that it uses caching extremely aggressively

Caching is fairly effective, but could improve

Coherent caching needed in high assurance systems but seems to be a much harder challenge