WebOS, PlanetLab, Starfish

Readings

Amin Vahat et al., "WebOS: Operating System Services for Wide Area Applications" (1998)

Adnan Agbaria and Roy Friedman, "Starfish: Fault-Tolerant Dynamic MPI Programs on Clusters of Workstations" (2003)

Larry Peterson et al., "Experiences Building PlanetLab" (2006)

Thomas Anderson and Timothy Roscoe, "Learning from PlanetLab" (2006)

---

WebOS

Key features

• High Availability
• Lower Latency
• Fault Tolerance

Consensus: WebOS isn't really an distributed OS

Main components

• Smart Client
• WebFS
• Global naming scheme based on URLs
• Process control system
• CRISIS authentication/authorization system (Certificates with ACLs)

Key ideas that were/were not not adopted from WebFS

Adopted:

• General idea of wide area dynamic distribution -> Akamai (but primarily for static content)
• Global naming using URLs

Not Adopted:

• CRISIS
• WebFS (Although WebDAV could be said to be related)
• Smart client (for web sites)

What are the pros and cons of using smart clients to do load balancing?

Pro:

• Distributes computation
• More flexible

Con:

• Vulnerable to Denial of Service or other forms of attacks
• Extra network overhead to locate a service

Starfish

Key features

MPI

• Fast message passing
• Allows the programmers more control

OCaml

Pros

• Recursive algorithms
• Uses Bytecodes - portable in theory
• Same code on heterogenous hardware
• Well known language

Cons

• Slow performance?
• Not entirely portable, maybe Just-In-Time?

Checkpointing

• Save / resume state
• Provides process migration
• Maybe better suited to be implemented in the application, not OS?

Management

• Had a Java interface which allowed any node to login and see full system status
• Used the distributed nature of daemons to communicate this information

Experiences Building PlanetLab

Key ideas

• Global platform to distribute/test network services
• Scale
• Should be able to monitor and stop disruptive traffic

Challenges

• Resource allocation is significant, especially in a distributed system such as PlanetLab
• Providing a global platform for long-term services
• Implementing a trust relationship between node owners and service developers (users)

PLC

Summary

• Fulfills the trust mechanism required for the network
• Acts as a middle man / mediator between node owners and users
• If someone breaks into the PLC however, entire system is compromised

Implementation

• Nodes are split into slices using VServers, lightweight process groups

Criticisms

• Bandwidth allocation was slice-based, not node based

Learning from PlanetLab

Centralized Trusts

Pro

• Easier to manage, one entity to trust

Con

• One point of control, no competition as to who to trust

Centralized resource control

Pro

• All resources are controlled by PLC

Con

• No incentive for users or administrators to conservce resources

Decentralized management

Pro

• Provide bare-bones management, try to foster competition between 3rd party services

Con

• No motivation to do so, it's hard work

Treat bandwidth as free

Pro

• Free bandwidth!

Con

• No incentive to conserve bandwidth

Provide only best-effort service

Pro

• No limit on the number of processes which can be run

Con

• Other processes may crowd out computation of other processes (cpu/disk hogging)

Linux is the execution environment

Pro

• Provides a familiar programming environment

Con

• Weak isolation between experiments
• No global allocation for resources
• Having a homogenous test-bed is poor for distributed experimentation

Evolve the API

Pro

• Adaptable API, ground-up

Con

• Never had a good API, inconsistent, ever changing, unstable programming environment

=== Focus on the machine room

Pro

• Allocate big machine here, other there, etc...

Con

• Bad for distributed OSes