Naming and Content Distribution

Transcription

1 Naming and Content Distribution To do q q q What s in a name Flat, structured, attribute-based naming and DNS Content distribution networks What's in a name? That which we call a rose By any other name would smell as sweet. W. Shakespeare, Romeo and Juliet (II, ii, 1-2)

2 Names, identifiers and addresses Names are used to denote entities in a system Hosts, printers, files, processes, users. To operate on an entity, e.g. print a file, we need to access it at an access point An entity can offer one or more access points (think phone #s) Access points are entities too Access points Entity (Lisa) An address is the name of an access point 742 Evergreen Terrace Springfield 2

3 Other names Pure names un-interpreted bit patterns Have no meaning, random strings (only good for comparison) No-pure ones contain info about the object (e.g., location, URL) (True) Identifiers names interpreted by programs Each identifier refers to at most one entity (no reusing) Each entity is referred to by at most one identifier E.g. phone is not, passport number maybe An identifier need not necessarily be a pure name, i.e., it may have content Alias a name defined to denote the same info as another name 3

4 Name service and names Name service Stores a collection of textual names, as bindings between names and attributes Name space All valid names recognized by a particular service A name is resolved when it is translated into data about the named object Why a separate service? Unification resources managed by different services using the same naming scheme Integration resources created in different administrative domain may eventually be shared 4

5 Names in distributed systems Flat naming Names chosen from a flat set of numeric or symbolic ids Must be globally unique e.g., broadcasting (ARP), hierarchical, DHTs Structured naming Have an internal structure that represent their position in a hierarchic name space Unique only within immediately containing level Each level resolved within the context of the next higher one Attribute-based naming Maybe easier to look up entities by attributes {(attribute, value)} Entities have a collection of (attribute, value) pairs 5

6 Structured, hierarchical naming DNS Host names malbec.cs.northwestern.edu Mnemonic nice on humans Variable length With little information about location IP address Numerical, nice on routers Fixed length Hierarchical, loosely based on location Naming on the Internet before DNS (1983) Each computer retrieved HOST.TXT from a computer at SRI (Menlo Park, CA) Single server issues A legacy a fossil host file still exist in most modern OS 6

7 Domain Name System (DNS) DNS (Paul Mockapetris, then at UC Irvine) A wide-area distributed database Goals: scalability, robustness, global scope, distributed updates, good performance Non-goals: No need for strong consistency Some uses Hostname to IP translation And reverse lookups as well Hostname aliasing other DNS names for a host Lookup domain s server by domain name RFC 1034,

8 Structured, hierarchical naming Hierarchical name space Names organized into name spaces Name space hierarchical as a rooted tree Domain name space partitioned organizationally and geographically Hierarchical name servers Root servers and a number of Top Level Domain servers NS hierarchy matches server hierarchy Authoritative DNS servers root Doing the translation Local DNS servers, near clients Resolver software running on clients com ar chicago eecs edu gov northwestern music 8

9 DNS root nameservers 13 root servers Each really a cluster relying on IP anycast 9

10 TLD, authoritative and local name servers Top-level Domains (TLD) Responsible for.org,.edu, country codes ar, ca, Authoritative DNS servers An organization s DNS server with authoritative info for that organization Maintained by the ISP or the organization itself Local name server Doesn t quite fit in the hierarchy Each ISP, company, university or department has 1+ When host make a DNS query, that s where it is sent Acts as proxy forwarding the query (resolving) 10

11 DNS resource records Each node in the hierarchy is a collection of resource records Examples of RR (name value ttl) SOA Holds info on the represented zone ( of sys admin, host where data on zone can be found, ) A IP address of the host this node represents MX Mail server to handle mail address to this node NS Name server that implement the represented zone CNAME Canonical name of the host (alias implemented by a node storing a CNAME record) HINFO Info on this host 11

12 Name resolution DNS in action To resolve a name we need a directory node; how do we actually find that (initial) node? Closure mechanism Many times implicit, e.g., in the Unix FS the i-node of the root directory is the first i-node in the FS Root Two type of queries Recursive NS response with answer or error Iterative NS may respond with referral (go talk to x) 12

13 A recursive DNS lookup.(root) authority edu: NS com: NS edu authority northwestern.edu: NS Talk to for edu Client Talk to for northwestern.edu Local nameserver northwestern.edu authority A

14 Recursive or iterative Interactive client drives the resolution Caching by clients only (a second client s resolution of the same name has to go through the same sequence ) Less burden on servers and more on the query initiator Potentially costly communication Recursive a name server passes result to next server Less burden on the one asking, higher demand on servers More effective caching Reduced communication costs Most root and TLD servers won t answer (shed load) 14

15 Scaling DNS Scalability though partitioning, replication and caching Tree sub-divides into zones beginning at the root Each zone could be 1+ domains and sub-domains Zone files the txt file that describes a zone Includes name and address for 2+ authoritative servers and for delegated subdomains Management parameters (e.g. caching) and RR Information in a zone is kept in 2+ name servers (redundancy) Any server can cache data from other servers If a non-authoritative server caches data, it notes the TTL 15

16 16

17 Trends and application need Some clear trends Growing number of networks Faster networks Growing availability and demand for content For applications, higher demand on performance and reliability Small degradation are expensive in lost revenue $2.8m/hour in 2009 damage reputation reduced productivity 17

18 Content delivery The common answer Replicate content around the world, closer to users Bring users to nearby content, nearby in a network sense A few ways to do this Content distribution networks deep into ISPs or bring ISPs to home Peer-to-peer Hybrid peer-assisted CDNs 18

19 Distributing content through CDNs Content provider determines which objects it wants the CDN to distribute tags and pushes content to CDN CDN replicates and pushes the content to its servers provides a mechanism for Replicating content on multiple servers in the Internet Letting clients pick the best servers to get the content from Mapping network proximity!= geographic proximity CDN replica CDN replica Client optimicdn 19

20 CDNs potential benefits Closeness to end users for performance and reliability Good scalability Avoid congestion and long latencies Redundancy for reliability and some resilience to DoS attacks Economies of scale Costly to maintain that many servers, control, replicate content, etc 20

21 Internet delivery challenges Peering point congestion Inefficient routing protocols Unreliable networks Inefficient communication protocols TCP can be a serious bottleneck to video delivery Scalability under and overprovisioning costs Application limitation and slow rate of change adoption IE6 still in use (<6%) Distance (server/user) RTT Typical packet loss Throughput 4GB DVD download time Local <100mi 1.6ms 0.6% 44 Mbps (high quality HDTV) 12min Regional 500-1,000mi 16ms 0.7% 4 Mbs (basic HDTV) 2.2hrs Cross-continent 48ms 1.0% 1 Mbps (SD TV) 8.2hrs Multi-continent~6,000mi 96ms 1.4% 0.4 Mbps (poor) 20hrs 21

22 Akamai as an example A deep into ISPs CDN Placing replica servers at ISP s POPs Distributed servers 100k servers, 1k of networks, 10s of countries Client requests >20m per second, 20-30% of all web traffic Customers Apple, BBC, FOX, MTV, NASA, 22

23 Components of a delivery network Edge servers/ replicas Origin End users Edge servers/ replicas Transport system Customers Communication and control system Mapping Data collection and analysis Management portal 23

24 CDN through an example When a browser is asked to get how does it know it should go to the CDN or get it from CNN? Users get an html document from this could be index.html index.html uses a modified URL for replicated content Example: If the jpeg files are what has been replicated then <img src= may be modified as follows: <img src= 24

25 CDN through an example What does this mean? <img src= host part: a73.g.akamai.net Akamai control part: /7/23 Content URL: /af/foo.jpg 25

26 CDN redirection The browser needs to resolve a73.g.akamai.net hostname for replicated content All DNS queries for g.akamai.net are sent to an authoritative DNS server for g.akamai.net Based on the IP address and information that it has about the Internet (called a map), the IP address of an Akamai regional server is returned to the requesting browser based on policy 26

27 CDN through an example Hierarchy of CDN DNS servers Multiple redirections to find nearby edge servers Customer DNS servers (3) Web replica servers (4) (2) (5) (6) Client gets CNAME entry with domain name in Akamai Client requests translation for cnn.com Local DNS (1) End user Client is given 2 web replica servers (fault tolerance) 27

28 CDN redirection Akamai IPs are cached at local DNS server Not always necessary to go to the root DNS server TTL associated with the IP address of an Akamai edge server is relatively small If content is not there Edge server gets it from others Or eventually from origin One tricky part, selecting the right edge server Want to spread load evenly Want minimal impact if server is added or removed 28

29 Mapping or server selection Picking a server Lowest load è To balance customer load Best performance è To improve client s experience Best on geography? RTT? Throughput? Load? Any server that is up è For reliability How to direct clients to the selected server As part of routing è Anycast As part of application è HTTP redirect 30X responses (301: moved permanently, 307: temporary redirect, ) As part of naming è DNS 29

30 CDN growth and impact Flattening of the Internet More content served from the edge Increase in peering Growth of IXPs More traffic at the edge, less in the core Changing economics Clients are happier with closer content So are CDNs And perhaps ISPs (reducing traffic from providers) CDN market worth Billion USD by 2022 Compound annual growth rate of 32.8% 30

31 Some interesting trends Content providers have CDNs If you are big enough it could be cheaper Tune caching to your particular service Can still rely on CDN services ISPs have CDNs Not just putting up pipes Reduce cross-isp traffic Hard to develop relationships with content providers Users have CDNs CoralCDN, BitTorrent, Hybrid solutions CDNs and ISPs Content providers and CDNs Peers and CDNs 31

32 CDNs models and markets Much more than Akamai Tons of commercial CDNs: Amazon CloudFront, BitGravity, CacheFly, CDNetworks, ChinaCache, CloudFare, Cotendo, Distil Networks, EdgeCast, Limelight, MaxCDN, Speedera, A few non-commercial ones: BootstrapCDN, CloudFare, Coral, Incapsula Some from telcos: AT&T, Bell, DT, Telecom, Telefonica, Level 3, 32

33 CDNs or P2P? P2P systems Cheap, easy to scale Security issues, potential low-quality, hard to find unpopular content, difficult accounting Infrastructure-based systems Expensive to setup and scale Akamai 137,000 servers in 87 countries (probably out of date) Can provide predictable QoS and reliable accounting Hybrid? Peer-assisted CDNs Deliver content by peers, with operation coordinated (and backstopped) by dedicated infrastructure Akamai s NetSession Operating commercially since 2010 True global coverage 239 countries in

34 CDNs or P2P? Both Hybrid? Peer-assisted CDNs Deliver content by peers, with operation coordinated (and backstopped) by dedicated infrastructure Akamai s NetSession Operating commercially since 2010 True global coverage 239 countries in 2013 Risks/Issues Need for revenue, unlike P2P No transparency users are aware of them Heterogeneity NATs and firewalls Impact to ISP change of traffic patterns 34

35 Some interesting trends Electricity costs of datacenters are high Estimated cost of Google in 2009 ~ over $38M/year And growing Systems growth outpacing energy efficiency gains Relative cost of electricity is growing Compared to hardware or bandwidth What s being done Energy efficient hardware, virtualization and consolidation, power off servers when possible, cheaper cooling Key observation: electricity prices vary on an hourly basis across markets 35

36 Price volatility across markets Locational pricing not well correlated CA-VA correlation California Virginia RT market price $/MWh Hourly variation Illinois peaks ~ $350/MWh negative prices day one day two day three time (hours) From A. Qureshi et al, SIGCOMM09 36

37 Price volatility across markets California has min. price 100 California Virginia RT market price $/MWh Virginia has min. price 0 day one day two day three time (hours) From A. Qureshi et al, SIGCOMM09 37

38 Some interesting trends Can you use request routing and replication to route away from high energy costs? [A. Qureshi et al, SIGCOMM09] 2% without increasing bw costs or worsening client performance Depends on the energy elasticity of clusters with full elasticity (an no bw constraints) over 30% Environmental impact, rather than energy? [P. Gao et al., SIGCOMM12] CO 2 emission of datacenters ~ Netherlands in 2008, reaching 2.6% of global total in 2020, > Germany Source (i.e., emission) changes per hour (termal generators on peak) Can you redirect traffic to a cleaner location? Gas 10% Washington Nuclear 9% Other Coal 6% 8% Hydro 67% Coal 37% Texas Other 8% Nuclear 10% Gas 45% Generator fuel type* 38

39 Summary Demand for content drives CDNs CDNs are interesting distributed systems Conceptually a virtual network Higher performance, reliability, security Works on the existing Internet as-is Alternatively, a clean slate re-design of the Internet could address the challenges Slow change due to sunk investment and entrenched adoption 39