Upload
elaine-fleming
View
212
Download
0
Embed Size (px)
Citation preview
1
Some Context for This Session…
Performance historically a concern for virtualized applications
By 2009, VMware (through vSphere) and hardware vendors nearly eliminated the costs of storage, network, and memory virtualization
With overheads near zero, new technologies in virtual deployments could sometimes beat physical counterparts
This session will focus on a diverse mix of extremelydemanding apps
Applications where we have proven performance with industry standard workloads and benchmarks
• Not just speeds and feeds
2
Paravirtualization
NPIV Support
TCP Segmentation Offload (TSO)
Jumbo Frames
Ubuntu64GB virtual RAM
256 GB of physical RAM
10 GigE
Infiniband
SATA devices
Windows Vista
Catalysts for Change - ESX 3.5 Platform Enhancements
Storage
Network
Virtual Machines
ESX Server
CPUMemory
Large memory pages
Performance Scale Compatibility
3
OS
APP
OS
APP
Storage
Networking
Virtual Machines
CPU
Memory
Hardware Scale Up 64 cores and 1 TB
physical RAM
Hardware Assist Purpose Built Scheduler
Lowest CPU overhead
Hardware Assist Page Sharing Ballooning
Maximum memory efficiency
VMXNET3VMDirectPath I/O
Wirespeed network access
Storage stack optimizationpvscsi
Less than 0.1 ms latency Over 350,000 IOPS
Virtual hardware scale out 8-way vSMP and 255 GB of
RAM per VMVM Scale Up
Current NEW
ESX
OS
APP
OS
APP
OS
APP
Catalysts for Change - ESX 4.0 Platform Enhancements
4
History Lesson on Application Performance
ESX Version
ESX 2 ESX 3
Ap
ps
Su
pp
ort
ed
100% ESX 3.5 ESX 4.0
Overhead
VM CPU
VM Memory
IO
• 30% - 60%
• 1 vCPU
• 3.6 GB
• 20% - 30%
• 2 vCPU
• 800 Mb
• 4 vCPU
• 64 GB
• 100,000 IOPS
• 9 Gb
• <2% - 10%
• 8 vCPU
• 255 GB
• >350,000 IOPS
• 30 Gb
• <10,000 IOPS
• 380 Mb
• 16 GB
• <10% - 20%
Source: VMware Capacity Planner analysis of > 700,000 servers in customer production environments
5
Catalysts for Change - Hardware Improvements
2006
2007
2010
2008
2009
AMD-V released
Intel VT-d released
Intel VT-d 10x fasterIntel EPT released
AMD-V 10x fasterAMD RVI released
Intel 4M L2 cache
AMD 4M L2 cache
Intel FlexPriority Released
6
Catalysts for Change - Software Scalability Limited
1990 1995 2000 2005 2010 20151
10
100
1000
Oracle SQL Server Exchange
Avg. Cust, Application Avg. Four Socket Web ServersYear
Nu
mb
er
of
Co
res
us
ed
Additional application scaling cost-prohibitive at some point
Virtualization provides a means to exploit the
hardware’s increasing parallelism
VMware ESX Scaling:
Keeping up with core counts
vSph
ere
X
vSph
ere
4
7
Catalysts for Change: New Virtualization-based Architectures
vSphere-based deployments have options unavailable to physical servers:
• On-loading multi-threaded IO drivers to efficiently use multiple cores
• Scale out on a single host
• Circumvents application scalability limitations
• Improves memory locality of reference and increases cache efficiency
• Hardware-accelerated network interrupt delivery
• Strict memory encapsulation into NUMA nodes
8
Catalysts for Change - Summary
VMware has made dramatic improvements in the performance of its virtualization platform
Hardware vendors have accelerated the efficiency of virtual workloads
vSphere provides flexibility that can allow administrators to circumvent application limitations
vSphere efficiently uses CPU in a way physical servers cannot
vSphere can meet and beat native application performance in many situations