XPDDS19: Core Scheduling in Xen - Jürgen Groß, SUSE
Art of Using Xen at Scale
1. Art
of
Using
Xen
at
Scale
Pradeep
Vincent
Principal
Engineer,
Amazon
Web
Services
2. Agenda
• Amazon
Web
Services
-‐
Intro
• Virtualiza@on
in
the
Cloud
• Scale
Impact
on
Virtualiza@on
• Xen
Enhancements..
• Future
Features
–
Some
Thoughts
4. Amazon
Web
Services
“..delivers
a
highly
scalable
cloud
compu@ng
plaIorm
with
high
availability
and
dependability,
and
the
flexibility
to
enable
customers
to
build
a
wide
range
of
applica@ons.”
5. Amazon
Web
Services
• Includes
more
than
20
Services
– Compute
Services
• EC2
– Storage
Services
• S3
• EBS
– Database
Services
– Content
Delivery
Services
– Messaging
Services
– And
many
more..
• More
Details
:
aws.amazon.com
6. EC2
Instance
Types
High
High
Cluster
Standard
Micro
Memory
Compute
Compute
Instances
Instances
Instances
Instances
Instances
8. Virtualiza@on
and
Cloud
• Virtualiza@on
a
key
enabler
for
Cloud
Compu@ng
• Virtualiza@on
Provides..
– Secure
Control
Plane
– Fine
grained
resource
alloca@on
– Cloud
Specific
Business
Logic
in
Dom-‐0
9. Virtualiza@on
in
the
Cloud
• Guest
OS(Kernel)
chosen
by
customer
– Large
and
ever
changing
set
• Aggressive
inges@on
of
Security
fixes
– Security
is
paramount
• Unique
Challenges
at
Scale..
15. VM-‐level
Fault
Isola@on
• Failure
related
to
one
VM
isolated
from
Peer
VMs
• Good
Hypervisor
Level
Fault
Isola@on
in
Xen
• End-‐to-‐End
Fault
Isola@on
Important
– Including
Dom-‐0
I/O
Stack
– Including
Control
Plane
(e.g.
Xenwatch)
22. Hardware
Enablement
• Xen
Enablement
bit
‘Slow’
for
some
hardware
• Faster
‘Produc@on’
quality
enablement
in
Xen
– Help
Cloud
adopt
new
hardware
faster
25. Interface
Compa@bility
• Applies
to
Many
Types
of
Interfaces
– Paravirtualized
Driver
Interface
• Blkfront
–
Blkback
• NeIront
–
Netbk
– Hypercall
Interface
– Control
Plane
Interface
(XM
vs
XL)
• Has
Been
Good
in
Xen
– Deserves
con@nued
focus
32. Mul@-‐@er
Scheduling
?
• Lots
of
use
cases
in
the
Cloud
• Classes
with
priori@za@on
– Beyond
‘CPU
Pools’
– Similar
to
Linux
Scheduler
?
– Real
Time
Scheduling/Advanced
Func@onality
• Preserve
exis@ng
Scheduler
Seman@cs
to
the
extent
possible
33. Dom-‐0
Resource
Accoun@ng
• CPU
consumed
in
Dom-‐0
for
each
VM
– Isn’t
accounted
against
the
VM
per
se
– Fairness
Issues
– Priori@za@on
Issues
34. Dom-‐0
Resource
Accoun@ng
• Possible
Solu@ons
– Group
Scheduling
in
Dom-‐0
–
Par@al
Solu@on
– Xen
and
Dom-‐0
Scheduler
Coopera@on
??
• Hypervisor
–
Dom-‐0
Interface
Blurred..
– Something
else
??
• Scheduler
deals
with
Threads..
– One
Thread
serving
Mul@ple
VMs
not
great
– ‘Thread
Aligned’
VM
Resource
Consump@on
Good
35. Key
Takeaways
• Virtualiza@on
Key
For
Cloud
• End-‐to-‐End
Fault
Isola@on
Important
• Performance
Isola@on
And
Performance
Consistency
Important
• Para-‐virtualized
I/O
Performance
Important
• Faster
‘Produc@on’
Quality
Hardware
Enablement
• Interface
Compa@bility
Cri@cal