OARtech Patrick W. Gilmore April 11, 2001

Agenda Who is Akamai? –Akamai’s Origins & Backgrounds –Content Distribution Network Edge Caches –Akamai Accelerated Network Program –Regions –Buddy System

Agenda Content Delivery –Object Delivery –Site Delivery –Video Streaming Mapping –DNS ä'kuhmy

Who Is Akamai?

Akamai’s Origins and Background Based in Cambridge, MA Founded by MIT research team –F. Thomson Leighton, Chief Scientist –Danny Lewin, CTO –Seeded in 1995 as an MIT research effort to improve Internet’s content distribution problem Team of 1,300 world class professionals Investors: Apple ($12M), Microsoft ($15M) and Cisco ($49M) IPO on 10/29 raised $250M 4th best in 1999

Current Akamai Customers

What is a Content Distribution Network? Akamai is a Content Distribution Network (CDN) That’s nice, but what is a CDN? –Three main components: Edge caches Content delivery Mapping

What is a CDN? Edge Caches –Working with ISPs and networks all over the world to install edge caches –More is better Content Delivery –Getting content to the edge –Includes object, videos, and whole web sites

What is a CDN? Mapping –Akamai’s Secret Sauce –Finding the closest edge server for each user –Network proximity, not necessarily the same as geographical proximity

How does Akamai’s CDN work? How does a Akamai’s CDN work? –All content must be hosted on an “origin server” accessible to all edge caches Working on Hierarchy, some customers using now –End users only speak to edge caches, not origin server, to get distributed content –Edge caches pull content once from origin server, then deliver content to many end users

How does Akamai’s CDN work? Why is this good? –Allows nearly infinite scalability (if deployed properly) –Good economies at large scales –Avoids congestion and long latency Speed of light issues, undersea fiber, etc. –Extremely reliable Mitigates some DoS attacks Massive redundancy

Edge Caches

Akamai Accelerated Network Program (AANP) What is the Akamai Accelerated Network Program? –Gives free edge caches to qualified networks Internet Service Providers Educational Institutions (usually Universities and K-12) Government networks Anyone else with “eyeballs”

AANP Benefits Performance –Content served locally, increasing performance Bandwidth Savings –Saves money on upstream, reduces congestion Reliability –A cache on a local LAN is more reliable than a web server on the Internet Interoperability –No routing or DNS changes, already using it today

AANP Benefits No Cost – FREE!! –Akamai pays for all equipment, shipping, etc. Easy to implement –Each Cache is just another another host on the LAN Akamai Support –24 x 7 NOC Co-Marketing –Logo use, press releases, trade shows, etc.

“Region” – set of edge caches All Regions –Intel based rack-mount PCs HTTP Regions –Linux based server –Proprietary HTTP server / cache Streaming Regions –Shrink wrapped software installed on our servers –Linux used for QT & Real –Win2K used for WMT

Regions Server Configuration –Dual Pentium III processors –Gigabyte of RAM –Two or four SCSI hard drives –Two 100BaseT network cards Additional Hardware –Ethernet switch(es) 100BaseT or Gigabit ethernet uplinks –Patch Cables

Buddy System Each server has two public IP addresses –“Service” address This is the address given to end users who want to retrieve content –“Physical” address This is the address used to manage and test the server If the physical address is not responding, the box is considered to be down

Buddy System Server Failure –Servers do fail occasionally –Each server has a “buddy” which is constantly trading hellos with the physical address –When a server stops responding to hellos, its buddy will respond to requests directed at the failed server’s service address –Users in the middle of a download may have to hit “reload” –No one else will notice any interruption

Content Delivery

End User Internet Host Server Last Mile Problem Possible bottlenecks on the Internet First Mile Problem Backbone Problem Peering Problem

Process Flow 1. User wants to download distributed web content 1 XYZ

2. User is directed through Akamai’s dynamic mapping to the “closest” edge cache Process Flow 1 2 XYZ

Process Flow 3. Edge cache searches local hard drive for content XYZ

Process Flow XYZ 3a 3b. If requested object is not on local hard drive, edge cache checks other edge caches in same region for object 3a

Process Flow 3b. If requested object is not cached or not fresh, edge cache sends an HTTP GET the origin server 1 2 3b XYZ 3 3a

3c. Origin server delivers object to edge cache over optimized connection Process Flow 1 2 3b XYZ 3 3c 3a

4. Edge server delivers content to end user Process Flow 1 2 3b XYZ 3 3c 3a 4

The Old Internet

The New Internet with Akamai

Case Study on Reliability and Scalability: The 2000 Election Crash Zone Without Akamai this site could not have served customers above their crash zone Customer Visits (Millions) Time

How a Non-Akamaized Website Works 1 User enters standard URL Customer Web Server Customer’s Web Server returns HTML with embedded URLs 2 Objects served with round trips across the Internet 4 User’s browser requests embedded objects from customer Web server 3 End User

Client’s Servers HTTP request user enters standard URL FreeFlow – Akamai’s Object Delivery Service HTML code contains Akamai URLs (ARL) Content Served Locally HTTP request for embedded content Akamai Server Example ARL: img src= a1000.g.akamai.net/…/

EdgeSuite – Akamai’s Site Delivery Service Customer CNAME’s (aliases) –Anyone looking up will be redirected to an Akamai hostname - “customer.d4p.net” No, I do not know why we use “d4p.net”. –customer.d4p.net is CNAME’d to aXXX.g.akamai.net –Standard Akamai mapping magic sends returns the closest edge server for aXXX.g.akamai.net

EdgeSuite – Akamai’s Site Delivery Service End user never communicates with origin server –Akamai retrieves content from private hostname Something like “origin.customer.com” –High reliability Thousands of servers backing each other up If one geographic area is disabled, no other area will be affected Mitigates some DoS attacks

EdgeSuite – Akamai’s Site Delivery Service Uncacheable content is tunneled back to origin –Can reduce need for tunneling with server-side scripts running on edge caches Persistent TCP connections increase performance –Helps with downloading of objects to end caches –Helps with tunneling to origin –(More on this later)

Video on Demand Akamai uses HTTP to transit content to edge caches –Allows lossless transport of content to edge –Origin server does not need streaming software, licenses, etc. –HTTP 1.1 byte-range request used to pull only required data User connects to edge cache with streaming protocol –Content is streamed in native format from cache to user –User experience is enhanced through “origin” file available on a “close” server

Live Akamai uses UDP to transit content to edge caches –Proprietary reflector network of servers –Akamai translates all streaming formats into UDP –Propagates live stream through reflector network Akamai Streaming Accomplishments: –First 1 Mbps stream live stream fed over public Internet –Single largest streaming event in Internet history Broke own record 3 times

Live “SteadyStream” User connects to edge cache with streaming protocol –Edge cache “subscribes” to reflector network to receive live streaming content –Three streams are delivered to each edge cache –Packet loss is eliminated through redundancy –Content is streamed from edge cache to user

Encoding Top-level refelectors Regions Dropped packets don’t degrade stream to regions X Lost connections don’t degrade stream to regions X X X X Entry Point Akamai SteadyStream SM technology can deliver streams reliably to the edge with effectively 0% packet loss. X

Live Pros: –Allows global distribution of real-time video –Users can connect anywhere and get good video quality over public Internet Cons: –Each region must have 3 users for effective bandwidth use –If streaming rate is > 1/3 access bandwidth to edge cache, congestion will occur –Reflector network uses bandwidth even if no users connect

QuickTime Streaming via HTTP Akamai can deliver Apple’s QuickTime files via HTTP –Delivered over HTTP caching network –HTTP network is more widely deployed than streaming networks –Packet loss is eliminated –Only good for VoD –More overhead than streaming protocols (uses TCP)

Live Streaming Over Satellite Network Satellite Uplink platform Satellite Uplink Facility Akamai Servers Akamai Switch Satellite DVB Receiver Akamai Streaming Servers To ISP Subscribers ISP/Network Datacenter Internet x IP Encoded Stream

Core Hierarchy Regions XYZ 1.User requests content and is mapped to optimal edge Akamai server

Core Hierarchy Regions XYZ 2. If content is not present in the region, it is requested from most optimal core region

Core Hierarchy Regions XYZ 3. Core region makes one request back to origin server

Core Hierarchy Regions XYZ 4. Core region can serve many edge regions with one request to origin server

Core Hierarch Features Reduces traffic back to origin server –Reduces infrastructure needs of customer –Provides best protection against flash crowds Especially important for large files ( e. g. Operating System updates or video files) Improved end-user response time –Core regions are well connected –Optimized connection speeds object delivery

Persistent TCP connections Avoids TCP slow-start and connection set up –Each TCP connection has to be set up with three packets –Once set up, a TCP connection starts at a low speed, and increases speed during connection Many web connections do not last long enough to reach top speed

Persistent TCP connections Reduces load on origin server –Maintain connections to a small set of Akamai servers instead of millions of end users –No new connections to be negotiated Setting up a new connection is significantly more CPU intensive than maintaining an existing connection

Mapping

Mapping algorithms –Heart of Akamai’s secret sauce –Directs end users to connect to nearest edge cache –Uses DNS to direct users to closest edge cache –Completely transparent to networks and end users – you are already using it today for the world’s most popular web sites

Mapping Mapping algorithms –Three main components to finding “closest” edge cache to end user from a Network point of view: Packet Loss Throughput Latency –Listed in order of importance (roughly)

Mapping Mapping algorithms –Mapping also takes into account edge cache performance Does a server have an object on its hard drive? –Uses consistent hashing algorithm (patent pending) Does the edge cache have CPU, RAM, bandwidth, etc. available to serve end-user?

Mapping Which is the best Akamai server to serve this end user?

Mapping Local Name Server –Computer asks local name server for IP address which matches hostname –Name will either be aXXX.g.akamai.net or an alias of that Name Resolution

Mapping Global Top Level Domain Server gTLD Servers Local Name Server –Local name server asks the global Top Level Domain (gTLD) servers who is authoritative for domain “akamai.net”

Mapping HLNS Root Name Servers Akamai High-Level name Servers 15 minutes Local Name Server –gTLD servers point local name server at Akamai High-Level name servers (HLNS) –Local name server requests IP address of aXXX.g.akamai.net from HLNS

Mapping HLNS Root Name Servers Akamai High-Level name Servers 15 minutes Local Name Server –HLNS looks at IP address of local name server and returns a sub- delegation for “g.akamai.net” pointing at a close Low Level name server (LLNS)

Mapping Sub-Delegation –Standard part of Domain Name System –Tells local name server “I do not know aXXX, but g.akamai.net is at this IP address” –Local name server automatically asks LLNS for IP address – transparent to end user

Mapping Mapping based on Local Name Server IP address –Note that the Akamai system has not seen the IP address of the end user –Leads to a small percentage of sub-optimally mapped end users –Working on ways around this Auto-Akamaizer Dynamic HTML in EdgeSuite Others

Mapping HLNS –HLNS has a “map” or matrix of IP blocks to LLNS Map weighs network proximity heavily, using geography only when all else is equal –HLNS “map” of network conditions is refreshed every minutes –Downed servers updated more frequently – approximately every minute Handled through Overflow Controller

Mapping Time To Live –Every time an IP address is returned to a name server, it is accompanied with a Time To Live (TTL) –Akamai returns the sub-delegation with a TTL of approximately 20 minutes

Mapping LLNS Root Name Servers Akamai High-Level DNS Servers 15 minutes Local Name Server 3-20 seconds Akamai Low-Level DNS Servers –LLNS picks local Akamai server which is not busy to serve content and returns that IP address

Mapping LLNS –Most edge caches do double duty as a LLNS –At least one LLNS outside local region is returned during sub-delegation to ensure end-user always gets a response –LLNS polls servers every 3 to 20 seconds for availability, load, etc. –At least two IP addresses are returned

Mapping TTL –The TTL of the final IP addresses is 20 seconds –This means if a machine goes down, the local name server will ask for a new IP address in 20 seconds –During that 20 seconds, the buddy will take over for the dead edge cache –If an entire region goes dies (rack loses power, etc.), the end user will experience a maximum of 20 seconds down time

Mapping Which is the best Akamai server to serve this end user? Root Name Servers Akamai High-Level DNS Servers 15 minutes Local Name Server 3-20 seconds Best Akamai Server for End User Akamai Low-Level DNS Servers Content Served Locally

Akamai \ ä'kuhmy \, adj. (Hawaiian) 1. Intelligent, clever. 2. “Cool.” Thank You !!!

Questions