Presentation slides at VEE 2014

1. Hwanju Kim12, Sangwook Kim1, Jinkyu Jeong1,
and Joonwon Lee1
Sungkyunkwan University1
University of Cambridge2
The 10th ACM SIGPLAN/SIGOPS International Conference on Virtual
Execution Environments (VEE)
Salt Lake City, Utah, March 1-2 2014
Virtual Asymmetric Multiprocessor
for Interactive Performance of
Consolidated Desktops

2. Virtual Desktop Infrastructure (VDI)
• Desktop provisioning
Dedicated workstations
VM VM
VM
VM
VM
- Resource underutilization
- High management cost
- High maintenance cost
- Energy wastage by idle desktops
- Low level of security
+ High resource utilization
+ Low management cost
(flexible HW/SW provisioning)
+ Low maintenance cost
(dynamic HW/SW upgrade)
+ Energy savings by consolidation
+ High level of security
(centralized data containment)
VM-based shared environments
2/20

3. Hardware
Virtual Machine Monitor (VMM)
Desktop Consolidation
• Distinctive workload characteristics
• High consolidation ratio
• 4:1~15:1 [VMware VDI], 6~8 per core [Botelho’08]
• Ever-increasing h/w parallelism (multi-core)
• Multi-layer mixed workloads
• Diverse user-dependent workloads
• Multi-tasking (interactive+background) in a consolidated VM
VM VM VM VM VM
VM VM VM VM
Mixed
Interactive
CPU-intensive Parallel
3/20
[VMware VDI] Enabling your end-to end virtualization solution. (http://www.vmware.com/solutions/partners/alliances/hp-vmware-customers.html)
[Botelho’08] Virtual machines per server, a viable metric for hardware selection? (http://itknowledgeexchange.techtarget.com/server-farm/virtual-
machines-per-server-a-viable-metric-for-hardware-selection/)
8 threads
(Core i7)
16 threads
(Nehalem EX)

4. Motivation
• Limited support for interactive performance
• Existing VMM schedulers give an illusion of
symmetric multiprocessor (SMP) to each VM
• Proportional share-based scheduler used for commodity
VMM (e.g., Xen, KVM, VMware)
pCPU pCPU pCPU pCPU
vCPU
vCPU
vCPU
vCPU
vCPU
vCPU
vCPU
vCPU
vCPU
vCPU
VM
Interactive Background
Time
shared
Virtual SMP (vSMP)
pCPU pCPU pCPU pCPU
vCPU
vCPU
vCPU
vCPU
vCPU
vCPU
vCPU
VMInteractive
Background
Virtual AMP (vAMP)
vCPU
Equally contended
regardless of
user interactions
Proposal
The size of vCPU =
The amount of CPU shares
Fast vCPUs Slow vCPUs
4/20

5. Goal
• Issue
• Accurately identifying a set of interactive vCPUs
• Design goals
• VMM-level approach
• Identifying interactive vCPUs unobtrusively using simple VMM
extension
• No share exchange
• Distributing CPU shares asymmetrically to sibling vCPUs within per-
VM share budget for performance isolation
• Optional guest OS extension
• Optionally installing lightweight guest OS extension for further
improvement of interactive performance
5/20

6. Workload Classification
• Previous methods
• Time-quanta based classification
[Kim et al., VEE’09]
• “Interactive workloads typically show short time quantum”
6/20
+ Clear classification between
I/O-bound and CPU-bound tasks
- Modern interactive workloads
show mixed behaviors
- Multithreaded CPU-bound job
shows short time quanta due to
inter-thread communication
[Kim et al., VEE’09] Task-aware virtual machine scheduling for I/O performance

7. Workload Classification
• Previous methods
• OS technique
• User I/O-driven IPC tracking [Zheng et al., SIGMETRICS’10]
+ Identifying a set of tasks involved in a user interaction (I/O)
- Relying on various OS-level IPC structures
 E.g., socket, pipe, signal
7/20
X server Terminal Firefox
IPC IPC
User I/O An interactive task group
[Zheng et al., SIGMETRICS’10] RSIO: automatic user interaction detection and scheduling

8. Workload Classification
• Challenges
• Time-quanta based scheme cannot accurately classify
modern desktop workloads
• VMM cannot access OS-internal IPC structure
• Key idea
• Tracking background tasks
• Identifying “background CPU load” before “user I/O”
• Interactive CPU load is typically initiated by user I/O
• VMM can unobtrusively monitor user I/O & per-task CPU load
8/20

9. • Proposed scheme
Workload Classification
9/20
VM
InteractiveBackground
QEMU/
SPICE
vCPU1 vCPU2
vCPU3 vCPU4
KVM Hypervisor
vAMP
Scheduler
Task Load Monitor
Background task information
Time
Interactive
phase
User input
T1
T2
T1
T2
CPUCPUCPUCPU
Background
Background
Task Load Monitor
• User I/O monitoring => I/O virtualization
• Per-task CPU load monitoring => task tracking technique
[Jones et al., USENIX’06]
[Jones et al., USENIX’06] Antfarm: Tracking processes in a virtual machine environment

10. Virtual Asymmetric Multiprocessor
• vAMP
• Dynamically adjusting CPU shares of a vCPU
according to its currently hosting task
1. Maintaining
per-task CPU load
during pre-I/O period
 Pre-I/O period is
set to shorter than
general user think time
(1 second by default)
2. Tagging tasks that
have generated
nontrivial CPU loads
as background tasks
 Threshold can be
set to filter daemon tasks
that possibly serve
interactive workloads
3. Dynamically adjusting
vCPU’s shares based on
weight ratio
(e.g., background :
non-background
= 1:5)
4. Providing vAMP
during an interactive
episode
 An interactive episode
is restarted when another
user I/O occurs or is
finished if maximum time is
elapsed without user I/O
10/20

11. Multimedia Workload Filtering
• Exceptional case
• Multimedia workloads (e.g., video playback)
• Can be misidentified as background workloads since it
continuously generate CPU load without user input
• Key observation
• Multimedia workloads generally accompany audio
output [Zheng et al., SIGMETRICS’10, Kim et al., MMSys’12]
• Solution
• Tracking tasks that access a virtual audio device
• Excluding audio access in an interrupt context
• Checking audio Interrupt Service Register (ISR)
11/20[Zheng et al., SIGMETRICS’10] RSIO: automatic user interaction detection and scheduling
[Kim et al., MMSys’12] Scheduler support for video-oriented multimedia on client-side virtualization

12. Limitation
• An intrinsic limitation of VMM-only approach
• A vAMP-oblivious OS scheduler
• Agnostic about underlying vAMP (i.e., all vCPUs are identical)
• Possibly multiplexing interactive and background tasks on the
same vCPU
• A slow vCPU has higher scheduling latency
• “Frequent multiplexing” might offset the benefit of vAMP
Example: A scheduling trace during Google Chrome launch
“vAMP might less effective if multiplexing frequently happens”
 Guest OS can be enlightened to mitigate the adverse effect of multiplexing
Background task Non-background task
12/20

13. Guest OS Extension
• Guest OS extension for vAMP
• OS enlightenment about vAMP
• To avoid ineffective multiplexing of interactive and
background tasks on the same vCPU  Isolation
• Design principles
• Keeping VMM OS-independent
• Optional extension for further enhancement of interactive
performance
• Keeping extension OS-independent
• No reliance on specific OS functionality
• Isolating tasks on separate CPUs is a general interface of
commodity OSes (e.g., modifying CPU affinity)
• Small kernel changes for low maintenance cost
13/20

14. Guest OS Extension
• Linux extension for vAMP
• User-level vAMP-daemon
• Isolating background tasks exposed by VMM from non-
background tasks
• Small kernel changes that expose background tasks to user
VM
vAMP scheduler
VMM
vCPU vCPU
Task load monitor
Background
tasks
T1, T2
vAMP-daemon
Kernel
User
Input
interface
Cpuset
interface
T1 T2
T3 T4
Procfs
interface
1. Event-
driven
2. Read
3. Isolate
Isolation procedure:
1. Initially dedicating nr_fast_vcpus to interactive
tasks (i.e., non-background tasks)
2. Periodically increasing nr_fast_vcpus when
fast vCPUs become fully utilized
(also periodically checking the end of an interactive
episode  stop isolation)
Default nr_fast_vcpus = 1 due to the low
thread-level parallelism of interactive workloads
[Blake et al., ISCA’10]
14/20
[Blake et al., ISCA’10] Evolution of thread-level parallelism in desktop applications

15. Experimental Setup
• S/W
• Linux KVM 3.0.0
• QEMU 1.0: handles I/O requests from guest OS
• H/W
• Intel Xeon X5550 2.67Ghz quad-core processor
• 8 pCPUs are available w/ hyperthreading enabled
• 8GB of DDR3 DRAM
• Measurement methodology
• Spiceplay: measures client-side performance
• Snapshot-based record/replay
• Robust replay for varying loads
• Similar to VNCPlay [Zeldovich et al., USENIX’05] and Deskbench [Rhee et al.,
IM’09]
• Extension on the SPICE remote desktop client
15/20[Zeldovich et al., USENIX’05] Interactive performance measurement with VNCplay
[Rhee et al., IM’09] DeskBench: Flexible virtual desktop benchmarking toolkit

16. Evaluation
• Application launch
• Background workload
• Data mining application (freqmine) with 8 threads
• Weight ratio (background : non-background)
• vAMP(L)=1:3, vAMP(M)=1:9, vAMP(H)=1:18
8-vCPU VM 8-vCPU VM
freqminefreqmine
App
launch
Remote
desktop
client
0.00
0.20
0.40
0.60
0.80
1.00
1.20
Impress Firefox Chrome Gimp
Normalizedaveragelaunchtime
Interactive applications
Baseline
vAMP(L)
vAMP(M)
vAMP(H)
vAMP(L) w/ Ext
vAMP(M) w/ Ext
vAMP(H) w/ Ext
vAMP improves launch performance by 7~40% High weight ratio is ineffective because of
negative effect of multiplexing
Guest OS extension achieves further improvement
of interactive performance by up to 70%
Why did Gimp show significant improvement
even without the guest OS extension?
8-pCPU
16/20

17. Evaluation
• Application launch
• Chrome vs. Gimp (without guest OS extension)
Chrome (Web browser)
Gimp (Image editing program)
 Many threads are cooperatively scheduled in a fine-grained manner
 A single thread dominantly involves computation with little communication
Background task Non-background task
Background task Non-background task
17/20

18. Evaluation
• Media player
• VLC media player
• 1920x800 HD video with 23.976 frames per second (FPS)
• Mult: multimedia workload filtering
8-vCPU VM 8-vCPU VM
freqminefreqmine
Media
player
8-pCPU
18/20
0
5
10
15
20
25
30
Averageframespersecond(FPS)
Baseline
vAMP(L) w/o Mult
vAMP(L)
vAMP(M)
vAMP(H)
vAMP(L) w/ Ext
vAMP(M) w/ Ext
vAMP(H) w/ Ext
Without multimedia workload filtering,
VLC is misidentified as a background task
vAMP improves playback quality by up to 22.3 FPS,
but high weight ratio still degrades the quality
Guest OS extension achieves 23.8 FPS

19. Conclusion
• vAMP
• Dynamically varying vCPU performance based on
their hosting workloads
• A feasible method of improving interactive performance
• Assisted by a simple guest OS extension
• Isolation of different types of workloads enhances the
effectiveness of vAMP
• Future work
• Collaboration of VMM and OSes for vAMP
• Standard & well-defined API
19/20

20. Thank You!
• Questions and comments
• Source code
• https://github.com/VirtualAMP
20/20