Technical Track presented by Roman Shaposhnik, Sr. Manager, Hadoop at Pivotal. OSv is the revolutionary new open source technology that combines the power of virtualization and micro-services architecture. This combination allows unmodified applications deployed in a virtualized environment to outperform bare-metal deployments. Yes. You've heard it right: for the first time ever we can stop asking the question of how much performance would I lose if I virtualize. OSv lets you ask a different question: how much would my application gain in performance if I virtualize it. This talk will start by looking into the architecture of OSv and the kind of optimizations it makes possible for native, unmodified applications. We will then focus on JVM-specific optimizations and specifically on speedups available to big data management distributed applications. Finally, we will look into the relationship between OSv and Docker and how that layering can help make OSv a secret sauce for turbo-charging Cloud Foundry application deployments.

1. © 2014

2. OSv: PROBABLY THE BEST OS FOR CLOUD
WORKLOADS YOU'VE NEVER HEARD OF
Roman Shaposhnik @rhatr, Pivotal Inc.

3. Why do we dig PaaS?

4. ‘cuz its awesome!

5. No, but seriously?
myApp $ cf push …
service #N
service #1
…
service #Z
service #M
…

6. What about stateful?
HDFS
Zookeeper
HBase
region
server
…
I am alive
Who Am I?
What do I do?
HBase
region
server

7. And when it fails?
HDFS
Zookeeper
HBase
region
server
…
Node died
You, pitch in!
Aye, aye captain

8. μservice-based deployment
HDFS
Zookeeper
HBase
region
server
…
I am alive
Who Am I?
What do I do?
HBase
region
server

9. Lets zoom in
…
service #N
service #1

10. Anatomy of a μservice
Hardware
“Stuff”
[Java] Virtual Machine
μservice code

11. How are we doing it today?
Hardware
pkg1
[Java] Virtual Machine
μservice code
pkgN………….
“OS”Linux kernel
Puppet, Chef
Huge VM image

12. Is there a better way?
vHardware
Hardware
[Java] Virtual Machine
libFS, libC, libJVM
μservice code
Application-specific
static linking
Tiny VM image AKA
unikernel
Hardware-assisted virtualization

13. What the heck is a FOOkernel?
• What OS design courses have taught us?
– microkernels vs. monolithic kernels
• What did they left behind?
– exokernels, nano
• What they should’ve taught us instead:
– unikernels, anykernels

14. Unikernels
• “Unikernels: library operating systems for the
cloud” came out in 2013
• A “library” operating system
• A kernel that can only support one process

15. Anykernels
• Programming discipline for kernel code reuse
• “The Design and Implementation of the
Anykernel and Rump Kernels” by A. Kantee
• Capabilities
– NetBSD filesystems as Linux processes
– User-space TCP/IP stack

16. OSv from Cloudius Systems
• A unikernel for “POSIX” and memory managed
platforms (JVM, Go, Lua)
• Anykernel’ish
– E.g. ZFS
• Runs on top of KVM, Xen, VirtualBox, VMWare
• Looks like an app to the host OS
• Small, fast and easy to manage at scale

17. OSv manifesto
• Run existing Linux applications
• Run existing Linux applications faster
• Make boot time ~= exec time
• Explore APIs beyond POSIX
• Leverage memory managed platforms (JVM, Go)
• Stay open

18. What’s inside?
single address space in “kernel mode”
“kernel threads”
“user threads”
diskZFS virtioC++ kernel
code
dynamic linker
libjvm.soifconfig.so
TCP/IP
iface

19. Anything it can’t do?
• A 100% replacement for a Linux kernel
– No fork()ing
• No process isolation
• The least amount of device drivers ever

20. Virtualization vs. performance
• Network-intensive apps:
– unmodified: 25% gain in throughput
47% decrease in latency
– non-POSIX APIs use for Memcached:
290% increase in performance
• Compute-intensive apps:
– YMMV

21. Van Jacabson’s net channels
socket
TCP
IP
iface
socket
TCP
IP
iface
lock
lock
lock
Traditional TCP/IP stack
app thread kernel (IRQ)
send/recv
socket
TCP
IP
iface
channel
classifier
iface
lock
OSv TCP/IP stack
app thread kernel (IRQ)
send/recv

22. Memory management in UNIX
OS Memory
Process Memory
JVM Heap
Process Memory
JVM Heap

23. Memory management in OSv
OS Memory
Process Memory
JVM Heap

24. JVM balooning (no more -Xmx)
JVM Heap
OS object

25. Turbo charging JVM GC
object 1 object 2

26. Turbo charging JVM GC
object 1 object 2

27. Turbo charging JVM GC
object 1 object 2
CPU MMU assisted tracking table

28. But what about Docker?
Jailed FS, net, etc.
Hardware
[Java] Virtual Machine
libFS, libC, libJVM
μservice code
Application-specific
static linking
Docker image
Common, shared kernel

29. Docker != LXC
• $ docker run roman/PetClinic
• Socially-driven image sharing
Ubuntu 14.04
Canonical
Tomcat
Pivotal
Go
Google
PetClinic
Roman
PetClinic
service #N
service #1
$ docker run

30. Why should it work this time?
• Unikernels/exokernels back in ’90
• JVM-on-bare-metal (Azul, BEA, etc.) back in ‘00
• Things they didn’t have back then
– HW-assisted virtualization (KVM, XEN, etc.)
– Elastic infrastructure oriented architectures
– CloudFoundry

31. Elastic, next generation datacenter
• Commodity, rack-provisioned Hardware
• Commodity, JeOS to get to Docker++
– CoreOS, SmartOS
• Docker++ as a common backed
• OSv (really KVM, XEN)
• “GitHub” for μservies images

32. Finally killing DevOps
• Ops (IT) maintains the bare OS
• Devs maintain the images

33. Finally killing DevOps
• Ops (IT) maintains the bare OS
• Devs maintain the images

34. Guinea pigs so far
• Cloudius
– Memcached
– Apache Cassandra
– Redis
• Pivotal
– Hadoop ecosystem

35. Where do we need your help?
• All of the above ;-)
• CloudFoundry integration
– $ cf push roman/petclinic ?
– Docker registry as a backend for BuildPacks ?
– Two-level scheduling for Hadoop ecosystem

36. Questions?
By @cloud_opinion
Imagine no platforms
I wonder if you can
No need for PAAS or IAAS
A brotherhood of bare metal
Imagine there is no VM
It's easy if you try
No host below us
Above us only apps