740 likes | 899 Views
By David Spurway, IBM Systems Architect . Bridging the cost gap between Power and x86. Agenda. Server Selection The initial cost problem All cores are not created equal Oracle DB example Why are non-virtualised servers under utilised? Virtualisation support Software costs
E N D
By David Spurway, IBM Systems Architect Bridging the cost gap between Power and x86
Agenda • Server Selection • The initial cost problem • All cores are not created equal • Oracle DB example • Why are non-virtualised servers under utilised? • Virtualisation support • Software costs • Adding redundant server • WAS example • Virtualisation built in or added on? • With or without limits? • Benefits of a big pool • Effect of workload spread • Software costs • Adding redundant server • Scaling up to larger customer levels • Combined Oracle and WAS example • Security • Other benefits • Summary
Many Factors Affect Choice Would you purchase a family car solely on one factor?
System z System x Power Selecting a Platform Time Horizon Scale ISV Support Non-Functional Requirements Deployment Model Technology Adoption Level Geographic Considerations Power, cooling, floor space constraints Platform Architecture Strategic Direction and Standards Politics Skills Cost Models
Click for table Single Server - Graph View Taken as single servers only, most of the IBM Power Systems servers are more expensive to purchase. The IBM Power 770 server is also larger and requires more power.
OLTP Performance results for Servers under consideration • The IBM Power Systems servers outperform the HP servers, particularly when looked at core by core. All cores are not created equal. IBM Power Systems can run more threads faster, which means workloads run faster and less cores are needed, which lowers costs
Matched to Benchmark Result - Graph View Click for table Working in performance, the difference in costs reduces, but the IBM Power Systems remain more expensive to purchase. Space and power become closer, and connectivity requirements are lowest with the IBM Power 770.
Typical server utilisation Typical UNIX or x86 serving or partition running a single operating environment is 10 - 20% utilised • Configuration planned for growth (10% added?) • Configuration planned for peaks (50% added?) • System waits for I/O and memory access even when it is working (20% unavailable?) What you pay for What you use Result is that 80% of the hardware, software, maintenance, floor space, and energy that you pay for, is unused unless virtualisation is used.
Oracle Certification For VMware and KVM • Running Oracle in a VMware ESX cluster you must license ALL of the cores in the cluster • Oracle DOES NOT recognise VMware as "hard partitioning" • http://blogs.gartner.com/chris-wolf/2010/11/10/oracle-broadens-x86-virtualisation-support-but-work-remains/ • Running Oracle in a VMware ESX cluster is not certified. If support is required for unknown problems then you must recreate the problem without VMware installed view Oracle Metalink document 249212.1 • Red Hat Enterprise Linux 5 integrates Kernel-based Virtual Machine (KVM) and ships Xen as the default hypervisor, so they are supported by Oracle under the Oracle Linux support program. However, Oracle does not support Oracle products on RHEL's KVM/Xen. • http://www.oracle.com/us/technologies/027617.pdf
IBM and Oracle Have a Long-Standing Relationship Sustaining relationship of 120K + clients • Oracle 22 years, PeopleSoft 20 years, JD Edwards 31 years, Siebel 10 years Coopetition is alive and well More than 120K joint technology clients • And more than 20,000 joint application clients Vibrant technology relationship • Sustained investment in skills and resources including dedicated international competency centres Market-leading services practice • IBM GBS is Oracle’s #1 SI partner (7,500 joint projects) with 5,000 people dedicated to Oracle Unrivalled client support process • Dedicated on-site resources and significant program investments Oracle Databases (along with most other Oracle products) are fully certified on IBM Power Systems, including the use of PowerVMvirtualisation, Micropartitioning, PowerHA and Live Partition Mobility (LPM certified for Single Instance DB only). http://www-03.ibm.com/support/techdocs/atsmastr.nsf/WebIndex/PRS3369
Effects of Utilisation - Graph View Click for table With utilisation effects added, the hardware costs become more comparable, and facilities requirements are much lower with the IBM Power 770. With IBM Power, virtualisation is built into the hardware and always on. This enables any workload to be virtualised, allowing high utilisation of resources. Less resources are therefore needed, reducing costs.
Add Oracle DB costs - Graph View Click for table Including Oracle DB costs makes the IBM Power Systems clearly more cost effective, with the IBM Power 770 showing the lowest costs and lowest facilities requirements. Some Enterprise Level Software has costs which are much higher than the hardware. IBM Power Systems can require far fewer processor cores, lowering these costs
Add Oracle DB costs (if ELA is in place) - Graph View Click for table Including Oracle DB costs makes the IBM Power Systems clearly more cost effective, with the IBM Power 770 showing the lowest costs and lowest facilities requirements. Some Enterprise Level Software has costs which are much higher than the hardware. IBM Power Systems can require far fewer processor cores, lowering these costs
Add Oracle RAC and Data Guard costs - Graph View Click for table As more software is considered in the model, the hardware costs become less prominent, and the IBM Power 770 remains the most cost effective. Some Enterprise Level Software has costs which are much higher than the hardware. IBM Power Systems can require far fewer processor cores, lowering these costs
Add Oracle RAC and Data Guard costs (if ELA is in place) - Graph View Click for table As more software is considered in the model, the hardware costs become less prominent, and the IBM Power 770 remains the most cost effective. Some Enterprise Level Software has costs which are much higher than the hardware. IBM Power Systems can require far fewer processor cores, lowering these costs
Add N+1 Oracle RAC Server costs - Graph View Click for table For this full Oracle RAC solution, the IBM Power 770 remains cost effective, and the IBM Power 740 is now the lowest cost solution. Some Enterprise Level Software has costs which are much higher than the hardware. IBM Power Systems can require far fewer processor cores, lowering these costs
Add N+1Oracle RAC Server costs (if ELA is in place) - Graph View Click for table For this full Oracle RAC solution, the IBM Power 770 remains cost effective, and the IBM Power 740 is now the lowest cost solution. Some Enterprise Level Software has costs which are much higher than the hardware. IBM Power Systems can require far fewer processor cores, lowering these costs
Add N+1 Oracle RAC Server costs at scale - Graph View Click for table For this full Oracle RAC solution, scaled up to the requirements of a large customer, the IBM Power 770 is the most cost effective, and has the lowest facilities requirements. Some Enterprise Level Software has costs which are much higher than the hardware. IBM Power Systems can require far fewer processor cores, lowering these costs
Add N+1 Oracle RAC Server costs at scale (if ELA is in place) - Graph View Click for table For this full Oracle RAC solution, scaled up to the requirements of a large customer, the IBM Power 770 is the most cost effective, and has the lowest facilities requirements. Some Enterprise Level Software has costs which are much higher than the hardware. IBM Power Systems can require far fewer processor cores, lowering these costs
PowerVM vs VMware IBM Power Systems always run in a virtualised mode, running in firmware • All IBM Power Systems benchmarks are published using PowerVM virtualisation • Unlike x86 virtualisation, there is no hidden performance cost for using virtualisation on IBM Power Systems • Other factors (Operating System, Database, etc.) had minimal effect in this case. 38% Overhead With VMware 4% Difference from other factors Source: http://www.sap.com/solutions/benchmark/sd2tier.epx
PowerVM vs KVM IBM Power Systems always run in a virtualised mode, running in firmware • All IBM Power Systems benchmarks are published using PowerVM virtualisation • Unlike x86 virtualisation, there is no hidden performance cost for using virtualisation on IBM Power Systems • Other factors (Operating System, Database, etc.) had minimal effect in this case. 40% Overhead With KVM 2% Difference from other factors Source: http://www.sap.com/solutions/benchmark/sd2tier.epx
Further VMware Limitations • Limitations to VMware 3.5 include: • 4 vCPUs per Virtual Machine • 4 Virtual NICs per Virtual Machine • 32 cores per server • 32 hyperthreaded logical processor per server • 256 GB RAM per Server • 10 GbE Unsupported • Above taken from this document - http://www.vmware.com/pdf/vi3_35/esx_3/r35/vi3_35_25_config_max.pdf • Planned DL580c G7 will not be fully supported (40 cores not supported, 20 cores with Hyperthreading not supported) • Limitation to VMware 4.0 include: • 8 vCPUs per Virtual Machine • 10 Virtual NICs per Virtual Machine • 64 Logical processors per host • Logical CPUs per host = CPU sockets x cores/socket x threads/core. Regardless of the host’s configuration of CPU sockets, cores/socket or threads per CPU core, the total number of logical CPUs (hardware threads) may not exceed this number. Logical CPUs in excess of this number are ignored. • 1 TB RAM per host • 4 10 GbE ports • Above taken from this document - http://www.vmware.com/pdf/vsphere4/r40/vsp_40_config_max.pdf • Planned 40 core DL580c G7 will not be supported with Hyperthreading • Hyperthreading can impact performance by ~30% - http://software.intel.com/en-us/articles/performance-insights-to-intel-hyper-threading-technology/
Scale to handle any workload Up to 64 cores, 256 threads, 4 TB of memory, and 600 PCI adapters with IBM Power 770 more possible with IBM Power 795 Virtualisation stack scalability – paravirtualisation features Scalable operating systems Strong ISV support and market share position Dedicated LPARs Shared Processor Pool Shared Sub-Pool Virtual I/O Server OS WPAR Virtual I/O Server OS OS LPAR LPAR LPAR LPAR LPAR LPAR PowerVM Hypervisor Click for vSphere 5 Side by side PowerVM Virtualisation Advantages Mix workloads with confidence • Hardware reliability • PowerVM is core firmware • Guaranteed capacity, multiple VIOS LPARs, dedicated devices if required • Live LPAR and AIX WPAR mobility • EAL4+ security rating Lower overall costs • High performance cores • Sub-capacity software licensing • Multiple processor pools • Dynamically add capacity only when needed – significant software savings • Workload partitions (WPAR) • Virtualised memory
Java Performance results for Servers under consideration • The IBM Power Systems servers outperform the HP servers, which continues across to the smaller IBM servers when compared to the smaller HP servers All cores are not created equal. IBM Power Systems can run more threads faster, which means workloads run faster and less cores are needed, which lowers costs
Adding Virtualisation to Benchmark Results - Graph View Click for table Rolling back to the performance results, and adding the 20% overhead for VMware has the IBM Power Systems as still the most expensive to purchase, but the facilities requirements become more competitive. With IBM Power Systems, virtualisation is built into the hardware, adds no overhead and is always on. Less resources are therefore needed, reducing costs.
Resource utilisation through virtualisation • Individual workloads tend to peak at different times, and need higher quantities of compute resources only when they are running at peak • At other times, resources may not be required to the same degree • Virtualisation can enable the use of these spare resources, driving higher utilisation of the physical resources • The larger the number of different workloads operating on the same physical server, the greater the benefit from resource sharing that can be observed • From detailed modelling of customer workloads, it has been observed that spreading workloads across additional physical servers can reduce the benefit from virtualisation • The total requirement for compute resource in this model increased by 15% with each additional server • Significant over allocation of resources is also possible through virtualisation • The sum of the peaks from each individual workload can commonly be 2 or 3 times the maximum peak on the virtualised server. Higher rates of over allocation are possible. • As CPU resources are used more effectively, additional memory can be needed • IBM Power Systems can accommodate more memory, and can use features such as Active Memory Sharing and Expansion to optimise the use of the installed resources.
Resource utilisation through virtualisationSingle Server Peak (minimum size of server) 59 cores for single server
Resource utilisation through virtualisationFirst server of two Peak (minimum size of first server) 25 cores
Resource utilisation through virtualisation Second server of two Peak (minimum size of second server) 43 cores
Results of splitting workloads from one server across two Peak (minimum size of server) 59 cores for single server = + Peak (minimum size of first server) 25 cores Peak (minimum size of second server) 43 cores A minimum of 68 cores needed in two server solution (an increase of ~15% over single server solution)
Resource utilisation through virtualisation First server of three Peak (minimum size of first server) 19 cores
Resource utilisation through virtualisation Second server of three Peak (minimum size of first server) 29 cores
Resource utilisation through virtualisation Third server of three Peak (minimum size of first server) 30 cores
Results of splitting workloads from one server across two or three Peak (minimum size of server) 59 cores = + Peak (minimum size of first server) 25 cores Peak (minimum size of second server) 43 cores A minimum of 68 cores needed in two server solution (an increase of ~15% over single server solution) + + Peak (minimum size of first server) 19 cores Peak (minimum size of first server) 29 cores Peak (minimum size of first server) 30 cores = A minimum of 78 cores needed in three server solution (a further increase of ~15% over two server solution)
Adding effect from multiple servers - Graph View Click for table Working in the effect of spreading the workloads over multiple servers shows the IBM Power 770 is competitive in the facilities requirements, but the IBM Power Systems still have higher purchase list price costs. IBM Power allows larger pools of virtualised resources. Spikes in workloads can be accommodated with fewer resources. Less resources are therefore needed, reducing costs.
Adding WebSphere Application Server costs - Graph View Click for table If the approximate costs for WAS over 3 years are added, then the IBM Power 740 is the most cost competitive, if running WAS alone as an application tier IBM Power Systems can still be competitive, even with less expensive software. Fewer servers are needed, reducing administration and facilities costs
Adding WAS costs with N+1 Server resilience - Graph View Click for table Adding in N+1 resilience makes the IBM Power 740 the most cost effective, if the servers are to be used purely as an application tier for WAS. On/Off Capacity on demand could be used to address the cost of the IBM Power 770 IBM Power Systems can still be competitive, even with less expensive software. Fewer servers are needed, reducing administration and facilities costs
Large customer scale for WAS costs with N+1 Server resilience - Graph View IBM Power 740 is the lowest cost. The IBM Power 770 does have compelling facilities requirements, as far fewer servers are needed. Click for table IBM Power Systems can still be competitive, even with less expensive software. Fewer servers are needed, reducing administration and facilities costs
Combined WAS and Oracle RAC at large customer scale - Graph View If IBM Power 770s are used for combined WAS and Oracle DB, particularly at the scale of a larger customer, they can have the best 3 year TCO. Also, the best facilities requirements are with the IBM Power 770s. Click for table At the scale larger customers operate at, both the software costs and the facilities requirements can be considerable. IBM Power servers can lower both, allowing high levels of expense to be avoided.
Administrative Costs 6% Infrastructure SW Licenses Support Costs 6% 10% Application Dev & SW Maintenance Support Costs 5% 25% DB+OS Licenses 3% Hardware Costs 8% Internal Implementation Costs 8% Implementation Costs 29% Cost Distribution in a sample ERP Implementation
The fastest patch time 11 minutes to apply a patch IBM Power Systems with AIX deliver 99.997% up time January 27, 2011 “For the third year in a row, IBM AIX Unix operating system (OS) running on the company’s Power System servers scored the highest reliability ratings among 19 different server OS platforms – including other Unix variants, Microsoft’s Windows Server, Linux distributions and Apple’s Mac OS X.” • Least amount of downtime • 15 minutes a year • 3.5x-4.5x better than Linux Source: ITIC 2009 Global Server Hardware & Server OS Reliability Survey Results, July 7, 2009
Click for 2011 & Total 2010 Reported Advisories Operating Systems Virtualisation Engines
Click for 2011 Solution Status 2010 Advisories Operating Systems Virtualisation Engines
Click for 2011 Criticality of 2010 Advisories Operating Systems Virtualisation Engines
Click for x86 Compare AIX and POWER7 RAS Features Virtualisation • PowerVM is core firmware • Thin bare metal Hypervisor • Device driver free Hypervisor • Redundant VIOS support • Dynamic LPAR operations • Separate HMC Users • Live partition mobility • HW enforced virtualisation support AIX • Integrated LVM and JFS • SMIT – reduce human errors • Hot AIX kernel patches • WPAR and WPAR mobility • App checkpoint/restart • Configurable error logs • Resource monitor & control • Role based access control • EAL 4+ security certification Virtual I/O Server Virtual I/O Server AIX AIX LPAR LPAR LPAR LPAR PowerVM General CPU Memory Network Disk General • First Failure Data Capture • Hot-node add/repair • Redundant clocks & service processors • Service proc failover • Concurrent firmware updates • CEC bus retry / recovery • Light path diagnostics CPU/Cache • Dynamic CPU deallocation • Processor instruction retry • Alternate processor recovery • Dynamic processor sparing • CPU CUoD • Processor contained checkstop • Dynamic cache deallocation and cache line delete Memory • DDR ECC Chipkill memory • Dynamic memory page deallocation • Storage protection keys • Memory bit steering / redundant memory • Dual sided DIMMs • Hardware memory scrubbing I/O • Redundant I/O links to I/O drawers • Independent PCI busses • Dynamic PCI bus slot deallocation • Hot swap disk, media, PCI adapters • Hot I/O drawer add AIX Security Expert Details AIX Encrypting Filesystem Details
Summary • All cores are not created equal • IBM Power can run more threads faster, which means workloads run faster and less cores are needed, which lowers costs • With IBM Power, virtualisation is built into the hardware, adds no overhead and is always on. • Less resources are therefore needed, reducing costs. • IBM Power allows larger pools of virtualised resources. • Spikes in workloads can be accommodated with fewer resources. Less resources are therefore needed, reducing costs. With the above factors being considered, the hardware costs for IBM Power can be comparable with the equivalent total HP hardware costs • At the scale larger customers operates at, both the software costs and the facilities requirements can be considerable • IBM Power servers can lower both, allowing high levels of expense to be avoided. • Some Enterprise level software has costs which are much higher than the hardware. • IBM Power can require far fewer processor cores, lowering these costs, resulting in a much lower cost of the solution • Other software has lower costs, but IBM Power can still be competitive • Fewer servers are needed, reducing administration and facilities costs • IBM Power and AIX then have a number of features that add value above Linux on x86 • More secure, less patching needed, can virtualise any workload, LPM, RAS, Active Memory Sharing and Expansion, WPARs, etc Taking in all these elements, IBM Power Systems offer solutions that can save money over x86 based solutions, and deliver higher levels of business value.
Conclusions and possible next steps • Hopefully my modelling and ideas have given some food for thought • But how the customer models costs and assigns value is far more important • Can we work with you on your models, working some of these ideas in, so IBM Power systems can compete effectively for workloads? Thank you! David Spurway – IBM Systems Architect Email: david.spurway@uk.ibm.com Phone: 07717 892 896