Defining and Improving
IT Utilisation Efficiency through
Holistic Data Centre Monitoring
Michael Rudgyard CEO
• A spin-out company of a well-established UK SI
• Technology was developed for High Performance Computing
– Management of HPC resources needs to be ‘system-wide’– Scalability (of both the architecture and the GUI) is paramount
• New company formed in March 2010
– Took on the product IP and existing HPC customer base – Notable investment from the UK Carbon Trust
• Currently in ‘semi-stealth’ mode
– Have developed new features for the Data Centre market
How Efficient is your
Data Centre ?
• Most new data centres are being designed against PUE targets
– For a given IT hardware capacity, PUE is a good planning metric – However, it is usually a poor operational metric• Most importantly: what if the servers are not doing any useful
work ??
– The data centre may still have a ‘good’ PUE, but it would be very inefficient by any business metric
• We really need a measure of IT Usage Effectiveness
– ie. how effective the power is being used to deliver necessary IT
• Unlike PUE, the concept of ITUE encompasses a family of
performance metrics
• Some metrics may provide useful generic ITUE measurement
– MIPS/watt or CPU Utilisation/watt (for compute bound tasks) – IOPS/watt or Bytes/watt (when I/O is predominant)• Some end-users may be interested in application-related metrics:
– Database transactions/watt– Page refresh/watt – Search/watt
• Some may be business related:
– £s of products sold / watt; or £s of products sold / integrated IT cost
• With few exceptions, the most successful methodology for
improving energy conservation across all sectors is:
– Step 1: Identify who/what is responsible for significant energy waste – Step 2: Drive behaviour to ‘encourage’ change
• What is the implication for the Data Centre ?
• Need to monitor and report ITUE metrics by customer,
department or end-user
– Who or what applications/service are the worst offenders ?
– Management can use data to drive better practice (charge-back ?)
• Efficient DCs will need to monitor & manage both IT and Facilities
systems in a coherent manner:
– Environmental systems (temperature, humidity, air-conditioning..) – Power (at the distribution board, rack PDU and server PSU level …) – IT equipment (using standard protocols such as IPMI and SNMP…) – Operating systems & Virtual Machines (integrating with IT systems) – ..and perhaps applications themselves
• Software tools will need to integrate with multiple systems from
multiple vendors (both hardware and software) in an agnostic
manner
• Optimised environmental management to improve PUE (& ITUE)
• Identification of unused, under-used, inefficient or over-spec’ed IT
equipment
• Using active power management during low utilisation periods
• Dynamic orchestration of virtual machines based on
environmental, power and IT usage constraints
• Non-trivial energy savings through simple changes (20-25%)
• Opportunity for very significant savings in most DCs (25-75%)
• Consolidation of Data Centres is already happening
– Driven by economies of scale and the ‘Cloud’
– The trend is only likely to accelerate…
• Conversely, as Data Centres become bigger, energy
management will become even more important
• The winners in the race for the clouds will be those
who can operate the most efficiently …
– .. but few know how efficient they are now !!
• The largest data centres are owned by a handful of IT
giants:
– Google, Amazon, Microsoft, Yahoo etc…
• These giants are very aware of Data Centre Efficiency
– Some have turned common perceptions on their heads
– Some even design their own servers
– All have developed their own systems and software
• Imagine a data-centre with 50 -100,000 servers (cf. Google)
– ie. 1,500-3,000 racks and a similar number of PDUs and sensors – and up to (say) 16 VMs per server• You might want to monitor (derive reports from & orchestrate…)
– 1,500-12,000 environmental sensors– 20-30 data-points per server (IPMI, Power) = 1-3M points
– 20-100 data-points per OS/VM (eg. SNMP, WMI) = 16-160M points – … as well as user and application data.
• That’s hell of a lot of information !
– But even scaling this back by an order of magnitude presents a challenge for software.
Things that won’t work:
• Using a ‘single-instance’ software architecture
– Information will need to be processed in a distributed manner
• Putting unrefined data in a standard SQL data-base
– or you’ll need another data-centre to store, process & retrieve the data !
• Expecting simple GUIs (eg. lists and trees) to be effective
– Visualisation becomes a key aspect to usabilityConcurrent Thinking’s
Products
‘Command & Control’ Architecture
• 1U appliance
• Collates information from concurrentCONTROL devices • Delivers highly polished, Web 2 GUI
Manage anywhere from mobile, Iphone, PDA, PC etc…
• Built to be a scalable interface
• ‘Zero’ U, low-power appliance
• Monitors data from devices associated with local ‘racks’
Power control and monitoring, Environmental information
• concurrentCOMMAND provides full system management GUI
• concurrentCONTROL devices act as slaves, and are designed to enable scalable and fault tolerant system monitoring and imaging
• Monitoring
– Power from power clamps, third party PDUs & PMBus PSUs – Environmental sensors: wired (5V) and wireless (866Mhz)
– Server hardware - IPMI, DCMI and Intel Node Manager support
– SNMP & WMI support for OS and VM monitoring; optional in-band ‘daemon’
• Reporting
– Power charge-back and ITUE metrics by group/customer/user/application – Scalable, ‘real-time’ data-centre views
– Extensive reporting of historical data
– Breach monitoring and reporting; Event data-base and visualisation
• Management
– Data Centre Inventory
– Power management (PDU and IPMI support) – Event scheduling
– Serial-over-LAN & SSH terminals
Visualisation of real-time metrics - data centre
view
Hardware repository: PDU association to
servers
• Integration with third party VMs (VMWare, KVM, Hyper-V ..)
– Dynamic orchestration of virtual machines