Whilst HPC systems form a relatively small part of most organisations’ IT infrastructure, they often demand the most attention. They are like the disruptive child at school who make up 5% – yet demand 30% of the resources. As a result, HPC systems are relatively unique and pose quite a challenge to most organisations who have them.
There is often a lot of misunderstanding and miscommunication when discussing the deployment of HPC systems and the impact they can have on the environment in which they are located. This misunderstanding, to a lesser or greater degree, will depend on the type of organisation that is deploying the system in terms of resources, experience and scale. It will also depend on who is responsible for the HPC itself.
For example, HPC systems often require a very different approach to power and cooling design plus, in most cases, are not considered business critical so they do not demand the same levels of resilience as other IT running business critical applications. To add to the complexity, there are two main types of HPC: Air-cooled and Water-cooled.
So what is high density HPC? How do you cool it? What are the different characteristics of these systems?
There is often confusion in the industry when people talk about high density and various methods to cool it. In most cases, HPC systems are cooled by air. This is when air is delivered to the front of the equipment drawn through the servers and hot air is exhausted at the back. There is further confusion when solutions are described as being water-cooled when in fact cooling units are ‘close coupled’ with the IT equipment itself. Examples would be in-row or front or rear door cooling systems. Whilst this uses water as in a conventional chilled water system it is actually based on an air-cooled solution for removing heat from the server and is not a water cooled HPC. Some larger and more extreme HPC systems use Direct Liquid Cooling (DLC) or on chip cooling where heat is removed direct from the chip using water. At Keysource, this is what we mean when we talk about water-cooled HPC.
High density IT can vary but tends to be described as a server rack of equipment which draws a lot of power, upwards of approx. 15kW per rack. In this range we see HPC systems that are air-cooled drawing 30kW and in some cases closer to 40kW per rack position. Some water-cooled (DLC) systems can go as high as 90kW + per rack.
When it comes to who has responsibility for the HPC, it is often not the business or the IT department that is deploying an HPC but a department which is focused on a particular project. The challenge then, for many organisations, is where to site these high density systems and who should specify the power and cooling solution to deal with it. If you put it into the organisation’s main computer room it will potentially have a negative impact on the other IT systems in the data centre by creating hot spots or disrupting airflow. This is because the traditional or older business critical IT systems do not have the same characteristics as the HPC.
An alternative is to deploy the HPC in a new environment which is designed specifically for the HPC. This might work for larger deployments where the scale justifies the investment but in most cases this is not so. We see this approach in organisations where one pot of funding is provided for an HPC project and the project owner uses some of that funding to build a new environment. Because the aim is to buy as much compute as possible, the environment is often value engineered or underspecified due to a lack of investment or expertise in providing appropriate power and cooling solutions. Therefore, this often leads to multiple data centres or computer rooms in a given location managed and run independently or a mixture of solutions in the same environment, and in many cases, these end up costing more and are not fit for purpose.
Therefore the challenge is simple. How can we create an environment which caters for mixed density IT systems be it HPC, networking or traditional business critical ICT?
If we put Disaster Recovery to one side as we accept the approach to this varies between organisations, it makes sense to house IT systems in the same physical location where you can control access, provide suitable environmental conditions and manage the infrastructure. However, most organisations have data centres that are existing and don’t have the luxury of time or money, or it is not practical to simply start from scratch.
At Keysource, the aim and role is to create an environment which allows the organisation to deploy what they want, where they want, when they want. It sounds simplistic and so it should be. We shouldn’t put constraints on the organisation in what they can deploy when it comes to IT and we shouldn’t have to re-engineer a solution every time a new piece of equipment needs to be deployed.
IT and HPC solutions change rapidly, therefore, organisations who can take advantage of the latest technology often gain a competitive advantage. In the world of research and HPC it can be very competitive and access to funding is difficult, so having an environment which is designed to accommodate this type of technology can be the difference between winning a project and not. On top of this, utilising dedicated and joined up tools, such as Data Centre Infrastructure Management software, give operators and planners real time information and insight to help manage capacity and plan deployment in a pro-active way.
Another often misunderstood concept is that of efficiency. Deploying high density IT that runs hot is an excellent way to deliver maximum efficiency in the data centre. This is caveated on the basis that the environment is suitably designed to cater for these higher densities. Assuming it is, we ideally want all of these systems located in the same place so we can scale the infrastructure and maximise the efficiency across all systems. With longer term power availability in some areas being questioned, along with the increasing cost of power, any measure which can reduce demand and on-going running costs helps with the business case for any project. It also frees up money for further development and investment in HPC, which is the main focus for these projects in any event.
Operating efficiency and effective use of capacity is also impacted by resilience. In a typical or conventional data centre, we work to design solutions which in most cases provide concurrent maintainability or even fault tolerance with fully redundant M&E solutions. However, most HPC systems do not require the same level of resilience. In that case, it is important to right size supporting infrastructure and not over engineer solutions when HPC is involved. Due to the density and potential total load of HPC, providing this resilience across M&E systems can significantly increase cost and complexity especially when you factor in Uninterruptible power supplies, and standby diesel generators. Working with the technical experts of the HPC and other stakeholders it is important to understand the impact of down time on all systems from processing, data storage, networking and other business applications so we can provide the right solution to meet the needs of the application.
A recent project for University of Leicester in the UK saw the University’s core data centre upgraded from a traditional legacy computer room which was sized for 60kW of IT and able to cater for just 5kW a rack. The upgrade delivered an increase in mains power and a redesign of the power and cooling system to cater for up to 300kW of IT. The project was driven by a new HPC system called ALICE which would require racks up to 30kW in any position and up to a maximum of 30 racks. The solution was to create complete separation of the hot and cold aisles but unlike many designs the cooling was not coupled to the rack. Instead Keysource’s ecofris design was installed which delivered the required airflow and heat rejection for the entire room. Using the sophisticated control system airflow and temperature are controlled based on the load and demand of the HPC. This ensures the system is running in optimum configuration all of the time, matching performance with demand. As the system is room-based, it also meant traditional and business critical IT can be deployed in the same footprint so it did not need a separate solution and is not affected by the HPC. Given that the solution was designed before the vendor of HPC was selected, this really was a case of wanting to be able to put what you want, where you want, when you want. Subsequently, phase 2 has seen the facility upgraded to double capacity to 300kW of IT load.
In conclusion, HPC is an excellent opportunity to innovate and in most cases provides a great opportunity to improve on current designs and efficiency. If all teams can work together and look to engineer solutions which offer maximum flexibility, we can see some real step changes in the attitude and approach to the deployment of HPC.