AIT

1. 1000’s computers

2. Power consumption in relation with compute load

3. Applications have cyclical demand curves

1. (50%,idle)

2. (70%,10%)

3. (90%,50%)

Most servers utilized 2. and 3.

1. Death of Michael Jackson on Twitter, Wikipedia, Google

2. Dynamically Pay by use

3. Amazon web services (AWS) elastic compute cloud (EC2) used by Farmville and Animoto

4. Statically provisioning for peak creates underutilization

5. Avoiding over-provisioning loses users

“Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction.”

Computing in the Internet

1. Resource pooling

2. Broad network access

3. On-demand self-service

4. Rapid elasticity

5. Measured service

6. No-need-to-know” due to APIs

Networking | Storage | ServerHW | Virtualization | Servers | Databases | Sequrity&Integration | Runtimes | Applications

1. Infrastructure as a service (IaaS)

2. Platform as a service (PaaS)

3. Software as a service (SaaS)

From database on

Only applications

Nothing all done by vendor

1. Private

2. Community

3. Public

4. Hybrid is a communication other clouds

1. Each receiver sets trigger with joint prefix k

2. Packet is delivered to node managing prefix

3. Postfix selection possible

ID can be far away and sender receiver close to each other

1. Close private IDs (randomly chosen and evaluated with metric as RTT) for private communication after initiation

2. Caches i3 node mapping specific ID

3. Install private triggers

1. Originally from active networking

2. Packet is sent to sequence (so-called stack) of IDS

3. Stack can be specified by receiver

4. If a trigger is reached replace ID with destination stack of trigger

“If nodes can fail, how can we store data persistently?”

Distributed data storage

1. Database management systems with key-value store

2. File systems with global file name-space

1. HBase 2. MongoDB and 3. Cassandra

1. Google GFS 2. Hadoop HDFS and 3. Kosmix KFS

1. By facebook

2. Open-source

3. Used by ebay Adobe IBM twitter netflix

4. Eventual weak consistency

1. Incremental partitioning feature

2. High data availability

3. Maintain membership among all nodes with Gossip

4. Tunable data consistency

5. Low down time during maintenance

1. Key-value store

2. Business // Website as twitter.com

3. Key // id or number as tweet id

4. Value // Information as tweet

1. Key-position store

2. Bloom filter

1. k hash functions

2. n items to be stored

3. m size of vector

1. Probabilistic compact way of representing items

2. Multiple hash functions in one vector

3. A bloom filter yes is probable false

4. A bloom filter no is definite

1. (1-(1-1/m)^kn)^k ~ (1-e^(-kn/m))^k

2. Optimal k~0,693*m/n

1. Column families

2. No schemas

3. Operations are get(key) and put(key, value)

4. Write heavy

1. Like SQL tables

2. Might be unstructured and client-specified

3. Might have index tables

1. Ring topology

2. Lightly loaded move position to alleviate highly nodes

Each data item at N (replication factor) nodes

1. Rack unaware is N-1 successive nodes after responsible one

2. Rack aware is coordinator telling nodes replica range

3. Data-center aware is like rack aware with coordinator chosen at data-center level instead of rack level

1. Determine node responsible of the data

2.1. Entry in disk commit log of sequential writes

2.2. Modify memtables and if full flush to disk

3. Responsible node updates all replica nodes // Wait for down nodes