C5

Question 1

distributed system

Answer 1

consists of multiple individual independent computers that communicate with each other
eg. cluster computer, interconnected network switchers and routers, peer-to-peer systems, cloud computing

goal: be resilient to failures

Question 2

horizontal scaling

Answer 2

adding more instances

• not as fast as vertical scaling
• if a full VM is not required, consider containers which can be spawned much quicker

Question 3

vertical scaling

Answer 3

increasing instance resources

• fast, but limited resources available in host computer
• several cloud providers don’t support online modification of a VM’s resources

Question 4

problems with scaling classic applications

Answer 4

session data: application layer often has state => make this tier stateless by pushing data into shared storage or cache

database tier: often hard to scale or unscalable (would need database cluster or distributed database)

Question 5

service-based architecture

Answer 5

develop multiple stand-alone services and build applications on top of this
- services interact over the network using pre-defined APIs
- no clear tier model, architecture is typically a graph of services
- all of these different services can be scaled

Question 6

elasticity

Answer 6

the ability to grow and shrink our pool of instances on demand

Question 7

MAPE for auto-scaling

Answer 7

Monitoring: monitor performance indicators such as request rate and CPU ultilization, and select right monitoring interval

Analysis: at what time to scale? scale proactively based on workload prediction

Planning: plan next scaling action, by what factor to scale, horizontal or vertical, what kind of resources

Execution: execute the plan through cloud API

Question 8

low-level metrics

Answer 8

• CPU ultilization
• memory statistics
• network bandwidth in/out

cloud provider customers won’t have access to low-level statistics of hypervisor

Question 9

high-level metrics

Answer 9

• request rate
• average response time
• session creation rate
• service time / request latency

not available to a cloud provider: application-dependent

Question 10

resource estimation

Answer 10

how many resources do we need?
need to determine the minimum amount of computing resources necessary to process a
workload and when and how to perform scaling

• rule-based: if metric reaches X, add N instances
• application profiling: determine saturation point of a single instance through benchmarking
• machine learning
• analytical modeling

Question 11

load balancers

Answer 11

assign clients to different servers to balance the load
- originally hardware products, later in software
- LBaaS