ebbrt-nodejs-slides.md

title	date	reveal_theme	reveal_transition	reveal_transitionSpeed	reveal_slideNumber
EbbRT+NodeJS (Unikernel 7/11/17)	2017-07-12	sky	slide	fast	false

Constructing Per-Application Operating Systems

James Cadden, PhD Candidate

Boston University

---

Elastic Building Block Runtime

EbbRT is a framework for building per-application library operating systems

==

Project Goal:

Reduce the effort to construct and maintain library operating systems without inhibiting specialization

==

==

==

==

---

Library Operating Systems

• Lightweight execution environment

• Deployed directly on (virtual) hardware interfaces

• Single-application, single address space

==

Application with LibOS and hypervisor

==

Decompose the LibOS into a set of system components...

==

...that are customizable, extensible, swappable, removable

==

Set chosen to fit the needs of a particular application

==

Statically compiled components combined at linktime

==

Application + OS compiled into single bootable image

---

EbbRT Native Runtime

• Per-core event-loop

• Non-preemptive execution / cooperative scheduling

• Identity mapped memory w/ large pages

==

Ebbs encapsulate major system component functionality

==

Building an Application

• Application source compiled using modified GNU toolchain

• x86_64-ebbrt build target

• Includes libstdc++, libc (newlib), C++/BOOST libraries

• Bootable .elf binary with kvm/qemu driver support

==

Elastic Building Blocks (Ebbs)

• C++14 objects with EbbRef type

auto myEbb=EbbRef<MyEbbType>();

myEbb->foo(); 

• Overload dereference operand to support specialized construction

==

Ebbs may have per-core representative instances

==

Instances are created lazily per-core

==

Instances are created lazily per-core

==

Instances are created lazily per-core

==

Or a single shared instance

==

EbbRT provides support for distributed Ebb invocation

==

Ebbs

• Statically compiled, chosen at link time

• Promote targeted customization

• Promote reuse of specialized implementation

---

OpenWhisk

==

• Invoke requests trigger a new action execution

• Consider client request latency (service times) and platform throughput

==

Invoker nodes deploys actions in isolated containers

==

Specialized invoker OS for deploying actions

==

• First-order design for runtime executions

• Simplifications user access, resource provisioning, isolation, devices,

---

EbbRT-NodeJS

==

EbbRT-NodeJS

==

JavaScript files arrive via network

==

JavaScript files arrive via network

==

Scripts are compiled and executed by V8

==

Dependencies decomposed into per-core Ebb instances

==

---

Extending the LibOS

==

libuv: multi-platform system support library (asynchronous I/O)

==

libuv: multi-platform system support library (asynchronous I/O)

---

Hosted Runtime

==

==

==

==

==

==

---

Next Steps

• Introduce multicore invoker to EbbRT-NodeJS

• Determine limits of existing libuv support

• Integrate EbbRT-Invoker into OpenWhisk

---

Thanks!

Questions?

---

Performance Results (OSDI)

==

V8 Benchmark Suite

==

EbbRT-Memcached Server Performance

==

Performance Advantages

• Event-driven multicore TCP/IP network stack

• Identity mapped memory w/ large pages

• Zero-copy I/O paths, asynchronous primatives

==

==

2x throughput at SLA

==

---

That's All Folks!