Unveiling The True Costs of HPC

Introduction

The ability to process massive amounts of data with great speed has earned HPC a place in a diverse range of industries, such as finance, healthcare, defense, meteorology, and many others. It can be an essential tool for a multitude of processes that revolve around discovery and innovation. Some examples would be 3D simulation, high definition film rendering, predictive analysis, molecular modeling, and sometimes just managing a sheer high volume of files, resulting from data that is generated quickly and in large quantities.  Yet, the impressive performance and other advantages also typically comes with a challenging topic – budgeting and financing.

Not only the hardware necessary for HPC is extremely expensive as it is, but the total cost – the true cost of HPC – extends far above and beyond the upfront costs. Our goal in this article is to go over the topic of HPC costs, including both surface and hidden costs, while also providing advice on how to reduce the total cost of HPC.

The Surface Cost of HPC

The surface cost of HPC implementation is the sum of various expenses that are obvious and apparent to any potential user. These are the costs that are considered initially when planning to implement an HPC solution. High-Performance Computing’s surface costs can be separated into five major categories: software, hardware, infrastructure, personnel and “other”.

Software licenses

There are three primary sub-categories of software that are considered a part of the HPC’s upfront cost. The first category is the payment for the operating systems – be it the enterprise Linux versions or completely custom operating systems made specifically to work with HPCs.

The second category covers everything a company might need to manage HPC hardware – resource monitoring tools, backup and recovery solutions, workload management software, job scheduling solutions, and so on.

Last, but not least, is the cost of the commercial software that can use the resources of HPC in order to perform specific workloads – data analysis, modeling, and simulation are just a small fraction of all the possibilities.

Hardware purchases

The most expensive element in the hardware department is also the most obvious element in the entire topic of “HPC costs” – the computing hardware that performs all the calculations and other tasks. Highly-tailored CPUs, GPUs, and memory configurations can be sold in server or cluster packages or separately.

Computing performance would not be nearly as effective without the storage capabilities to support this kind of read-and-write operation speed. As such, there is also a necessity to use high-speed storage solutions – not only SSDs (faster but less storage) or HDDs (significantly slower with higher capacity) but also highly customized storage systems capable of handling HPC workloads, such as parallel file systems.

Most existing infrastructures and systems already operate using some form of a wireless connection. The hardware in question also has to keep up with the performance of the HPC frameworks, which implies expensive Ethernet switches and routers capable of handling high bandwidth, and plenty of other high-speed hardware for networking purposes.

Practically every customer-oriented computing system produces heat when performing calculations or other operations. HPCs are no exception to this rule, and the cooling systems installed in this hardware should be capable of handling extremely high temperatures to make sure that the expensive hardware would not melt itself down under heavy workloads. Two of the most popular options so far are liquid cooling systems and air cooling systems, and the choice between the two greatly depends on the specific circumstances of each HPC instance.

Infrastructure configuration

Setting up the hardware and providing power to it are both important parts of the initial HPC setup. First of all, high-performance calculations mean that there is a lot of power needed on a constant basis. Investing in reliable and scalable power supplies is paramount in the context of HPC infrastructures.

The sheer complexity of this hardware also means that the setup and initial customization would have to be performed by a professional whose services also cost money. Additionally, the total cost of storing all of this hardware in a dedicated facility with all the necessary elements, such as electricity and physical security, is also a significant portion of the total upfront cost for HPCs.

Workforce expenses

A relatively small portion of the expenses also goes towards training. Not only is it likely necessary to train your existing employees in regards to how the HPC operates, but there are also perhaps additional people that would have to be hired just to maintain the working conditions of an HPC unit – be it HPC specialists, system administrators, and so on.

The Hidden Cost of HPC

As the title suggests, hidden costs in the context of HPC represent potential expenses that might have been overlooked during the planning and budgeting phases. The combination of all hidden costs is often greater than the total sum of surface costs, which is why we are going over this topic in the first place.

The total number of cost groups here is going to be a lot higher since it is difficult to position a lot of these costs any other way than completely separate from the rest. Nevertheless, we are going to go over six primary cost groups: operational expenses, facility expansions, data security, downtime, data management, and compliance.

Operational expenses

Regular maintenance on both the hardware and the software sides is essential in order for HPCs to keep running at their best possible efficiency. The electricity would be another notable option here, but we have already mentioned it before as it is a rather obvious choice. Applying various patches and upgrades to the existing software is an important part of any HPC system to improve performance, boost security, and so on.

Facility expansions

The aforementioned upgrade policy does not just apply to the software side of the tool, but it also should be performed on the hardware side. Modifying existing HPC hardware in order to achieve better power distribution, higher performance, or more efficient cooling systems is also considered a hidden cost in the long run. There might also come a time when the current HPC storage accommodations are not going to be enough for the current state of the tool. In that case, facility expansion or new space leasing is also going to belong to this category of hidden expenses.

Data security

Robust security measures are almost certainly a necessity, considering HPC’s use in very specific security-conscious fields such as research and development. The task becomes even harder considering the sheer size of an average HPC environment. A multi-pronged approach with customizable data security policies is a must in most situations. The majority of HPC and supercomputing applications are used for highly valuable and often highly sensitive data and applications. Therefore, security is likely a critical factor for most HPC organizations.

Downtime

Every single instance of downtime, be it expected or unexpected, gathers losses in both productivity and potential revenue. Avoiding these situations as much as possible is the priority of any modern infrastructure, and HPC is no exception. Fine-tuning both the hardware and software sides of the HPC is also an ongoing process that is practically endless in its nature. There is always going to be something that can be done to improve the results, be it a new piece of hardware, a new software update, etc.

Data management

Dealing with large data sets is a significant challenge in any system. Transferring large data masses to a different location, be it for the sake of preservation or for any other reason, is a very delicate process that is also quite expensive due to the fact that the receiving side of the transfer should have the capacity and the performance to handle it. Both cloud-based data storage solutions and remote data centers have their own shortcomings (for example, some cloud providers even have egress charges).

Additionally, not all storage in the HPC has to be approached the same way. There is plenty of potential to use various forms of long-term storage along with the archiving processes, which can save storage space, improve performance, and gain a host of other potential advantages.

Compliance

There are many different regulations and industry standards that affect HPCs in some way. Most of the existing regulations imply extremely steep consequences for not following their guidelines, from monetary fines to lawsuits and sheer reputational damage. Compliance is paramount, even if it can be challenging to implement in environments as vast as the HPCs are.

Most of these costs are very difficult to predict or even offer a specific example of. However, it would be wise to mention that every single element of the pricing structure can differ in terms of both quality and price. Moreover, it is only possible to use the higher “grade” of specific pricing elements in conjunction with several other points on the list. For example, choosing to invest in more powerful hardware such as CPUs or GPUs would offer practically no benefits unless more investments are made into supplementary hardware – high-performance network environments, fast storage solutions, and so on.

Cost-reduction techniques for High-Performance Computing environments

HPC management can be a rather expensive and challenging task. While the specific nuances of each situation might differ from one case to another, we can offer ten different recommendations in terms of how a company can improve the cost-related situation in their HPC environment:

1. Workload consolidation can be used to improve the overall resource utilization of the environment by allowing more than one application to run on the same hardware when applicable. The results of this change are a lower total number of servers required, lower operational costs for the HPC, and better resource optimization overall.
2. Vendor contract renegotiations every once in a while can offer substantial advantages to their users when done right. Since it is highly likely that an average HPC environment would have contracts with not just software but also hardware firms and even service vendors, renegotiating has plenty of potential to offer better terms, lower pricing, and other advantages. Some of the most possible ways to do so are multi-year agreements, bulk purchasing, and even competition leveraging, when possible.
3. Total electricity consumption can be reduced by implementing better cooling solutions and more energy-efficient hardware and software. It should even be possible to use advanced power management in order to reduce power consumption of the entire HPC infrastructure outside of the peak workload hours.
4. Researching the existing open-source alternatives for the HPC software can lead to some of the licensing costs being alleviated. However, it is worth noting that many of the open-source solutions are notorious for being difficult to master, even compared with high-level paid enterprise software.
5. Containerization and virtualization can also be implemented in order to optimize resource usage within the HPC environment. Technologies such as Kubernetes, VMware, Docker, and several others can be used to reduce the number of physical servers required to run all of the necessary software (by allowing multiple software instances to run on the same server), offering improvements in terms of the hardware costs, resource allocation, and so on.
6. Staff training programs, while expensive, can drastically improve the HPC management capabilities of the company, leading to a lower number of errors caused by the human factor (which is usually a lot more than the cost of the staff training efforts).
7. Cloud bursting is a relatively new approach to cloud storage implementation in environments similar to HPC. It uses cloud resources that are only added to the infrastructure during known peak workload periods, which drastically reduces its total cost and makes the entire system a lot more cost-efficient.
8. Tiered storage solutions and other means of improving data management efficiency can also serve as a rather substantial storage expenses optimization factor. The idea behind it is relatively simple – the most commonly used data is kept at the most high-speed hardware, while the less critical information can be stored on a slower and cheaper storage.
9. The allocation of HPC resources can be further improved by adopting comprehensive job management solutions and workload schedulers (Grid Engine, PBS, SLURM). The final result is going to differ significantly in every single situation, but general throughput improvements and idle time reduction are practically guaranteed to the majority of HPC users.
10. Ongoing optimization for the existing HPC environments is made possible by conducting regular audits. Being able to identify inefficiencies and other lackluster elements of the system makes it possible to improve the performance of the system on a continuous basis and even remove resources that are considered redundant and no longer necessary for the system to function.

How Bacula Enterprise can assist with HPC cost reduction

HPC environments are large and often challenging to manage – and yet, they still need all of the basic security measures, including something as simple as creating data backups. Luckily, enterprise-grade solutions such as Bacula Enterprise are capable of offering data backups to HPC environments and many other subsequent capabilities.

Bacula’s robust and highly scalable solution can offer impressive data protection capabilities, flexible architecture, and efficient storage space utilization in a single package. It uses a flexible licensing model that is not tied to the storage capacity, making it a massive advantage to the HPC environments (since they often handle tremendous data volumes).

Bacula also excels in disaster recovery, does not tie its clients to a short list of supported vendors, and can even offer some of the advantages of an open-source nature (since it was created as an extension of a free and open-source solution to begin with). Bacula is used by many of the largest research centers in the world, including NASA, the National Laboratories in the USA, and some of Europe’s leading research organizations.

Sustainability considerations and Bacula Enterprise

Sustainability is rapidly becoming increasingly important in HPC, supercomputing and large data centers. Bacula Enterprise is also especially aware of the modern environmentally-conscious landscape, offering sustainability as a key talking point in a wide variety of its features: infrastructure utilization improvements, support for many environment types, advanced storage-saving techniques to reduce storage requirements, economical management of billions of files, and the overall focus on reducing energy consumption for both backup and restoration processes. Bacula exceptional security levels, compared to other backup vendors, are also key for sustainability and business continuity. Its open-source background means its software code was created in a far more sustainable manner, and its main offices are based in Switzerland, where low-carbon power production is far ahead of that of the other home countries of Bacula’s peers.

That way, Bacula’s solution manages to lower the carbon footprint of the data protection operations, contributing significantly to the popularization of sustainable data management practices, extending the life of the existing or legacy hardware to minimize e-waste, and so on. Please contact Bacula for more information, and its white paper on its especially high sustainability levels.

Conclusion

High-Performance Computing environments can handle workloads of tremendous sizes and are pretty important in many industries, especially when it comes to performing various calculations, estimates, and data-driven decisions. At the same time, setting up a single HPC environment is a massive undertaking in terms of both time and resources.

Surprisingly enough, not all organizations recognize that managing an HPC environment can turn out to be far more expensive than initially purchasing it and setting it up. This is why we have separated the article into surface costs and hidden costs, offering multiple categories and sub-categories as the means of explaining the topic.

Cost explanation was not the only purpose of this article, though. We have also shared tactics and recommendations that could lead to general budget improvements, be it virtualization, workload consolidation, contract renegotiation, and more. The purpose of the third-party backup software was also covered in detail, with Bacula Enterprise being used as an example of how much a highly scalable backup and recovery solution can be critical to such a complex and multifaceted environment.

HPC can be a very valuable tool for many different industries in this day and age. Knowing about the total price of such an environment and which key elements contribute to it should serve as a valuable source of information when evaluating the total cost of the implementation in both the short- and the long-term.