Paid Feature Nobody really expects their data centre workloads to become less demanding over time. But for more than a year and a half, infrastructure managers have had to facilitate not just the increasing take-up of advanced analytics and AI, with their ferocious appetite for data. They’ve also had to deal with the disruption caused by the COVID pandemic, which has meant retooling to support remote workers, often through the rollout of memory-hungry virtual desktop infrastructure platforms.
And, if they have time, they have to support a myriad of other efforts, from digital transformation and cloud native, to IoT, to 5G and everything else in between.
It’s no surprise that many organisations find their servers running hot as a result – literally in some cases. It’s not hard to work out why. Your management software should point you to where the bottlenecks are. Something as straightforward as opening Windows Resource Manager can show you how choked your memory is – and why – or highlight how your storage is straining.
So, what can you do? You might consider shifting to the cloud, but there’s a good chance you might find that the cookie-cutter CPU-memory-storage configurations on offer do not suit your workloads. And that’s before you take into account the migration headaches and hidden costs a cloud shift can mean – or the compliance issues that can arise.
It may make far more sense to consider new servers, either in addition to your existing fleet or as a fully-fledged rip and replace. More CPUs with more cores, and more memory, and more storage, all safely stamped with the vendor’s seal of approval.
It’s no surprise that IDC found that worldwide server shipments were up 8.3 per cent to nearly 2.8 million units in the first quarter of this year. That growth partly reflected the fact that Q1 2020 was when the pandemic first impacted the economy, said IDC, but also reflected “increased investments targeting the modernisation of business applications, data centre infrastructure, and IT operations.”
But simply adding more tin – or replacing older equipment – is not an option for everyone. Even if you do get sign-off from your CFO, ripping and replacing servers is never a trivial undertaking. Simply speccing out what you need is time consuming, never mind the time and energy to physically decommission older tin and replace it with new kit. Plus, even if you can live with the environmental impact of junking perfectly serviceable kit, some of your colleagues might raise some noisy objections.
And the fact is, you’ve probably just kicked the can down the road. Because your new servers might simply have reproduced the same bottlenecks and they’re just there, waiting to reveal themselves over time.
The eternal trinity
This is because although enterprise workloads and architectures might have changed dramatically over the decades, compute still rests on the eternal trinity of CPU, memory and storage. In recent years, CPUs have acquired ever more cores, massively increasing their potential to do work. But this depends on the CPU cores having sufficient data to work with in the first place.
Intel’s Xeon processors, just two years old, range from four cores to 56, and support from 1TB of DDR memory to 4TB of DDR per socket in the top end models. Yet that processing horsepower is often shamefully underutilised because the system is simply not supplying them with data quickly enough. IO constraints play a part, but standard server configurations will often skimp on comparatively expensive DRAM, and further down the data path, bulk storage. (Let’s assume that your storage is SSD-based, rather than relying on hard disk drives.)
Again, opening the Windows resource tab might show the disk and memory panels soaked in green and blue, while the CPU appears to be virtually flatlining. So, it’s logical that simply boosting DRAM and storage can give a massive performance boost to even only slightly used systems – it’s something we all know, but don’t always consider.
Upgrading DRAM means more data is primed and ready to load into the CPU, speeding up operations and ensuring you’re getting the most out of your very expensive server CPUs, whether that’s loading up datasets for a heavy weight analytics exercise or streaming massive video files.
Of course, you want to ensure that DRAM is also fully utilised too. So, faster storage, in the shape of SSDs, will ensure faster writes to and from DRAM. This can be perked-up even faster if you plump for NVMe storage.
Simply increasing storage capacity can boost performance. The more free space in an SSD, the faster its overall performance. The closer it gets to full capacity, the higher the performance penalty. And if your servers are still using hard disks switching to an SSD will lead to dramatically faster application responses.
But you can tailor things even further. SSDs are not all created equal. Depending on your workload, you can opt for devices geared towards raw capacity or for read or write access. When it comes to AI or analytics workloads, the training stage for an algorithm might call for very large capacity devices. Inference or analysis tasks, by contrast, might favour comparatively smaller devices that might offer an edge in data access speeds. Micron, for example, categorises its SSD drives as read-centric, write-centric or mixed-used, meaning that customers can match their storage to specific workloads – and avoid paying for features they don’t actually need.
So a judicious upgrade of memory, or storage, or even better, both, will not just prolong the life of your existing server investments – and associated infrastructure - but deliver a significant performance boost. And this can be done in lockstep, or as separate projects.
For UK organisations, there’s another factor to take into account. From April 1, 2021 to the end of March 2023, UK companies can claim a 130 per cent capital allowances on qualifying plant and machinery investments, thanks to the “super deduction” policy announced by Chancellor Rishi Sunak back in March.
This includes computer equipment and servers, meaning, in effect, that for every pound a company invests, its taxes are cut by 25p. According to the UK government’s super deduction fact sheet, a company making £1m of qualifying expenditure can deduct £1.3m in computing its taxable profits, saving the company £247,000 on its corporation tax bill.
The government’s aim is to stimulate the economy in the wake of the disruption caused by Covid, as well as to incentivise business investment by UK businesses and close the UK’s lingering productivity gap.
So, what’s your next step? It’s probably not going to be to reach for your server supplier’s parts catalogue. OEM parts – i.e. those sourced from the server vendor - are generally priced at a substantial premium. Calculations by Micron earlier this year suggested that fully loading an industry standard with OEM labelled memory and storage will come in at almost three times the cost of the Micron-sourced equivalent.
It’s worth bearing in mind that there aren’t exactly lots of NAND manufacturers in the world, meaning an OEM product is simply a rebranded but certified part. While there may be a certain sense of security in buying a manufacturer certified part, it’s also worth considering that certification takes time, meaning your manufacturer branded part may not be the most up to date spec available.
A more tangible concern might be that using a third party product could cause problems with warranties or service contracts. Some vendor’s warranties will say that use of unsupported parts will void a warranty, or exclude any damage caused as a result of a botched upgrade. The short answer is the situation will vary from customer to customer and vendor to vendor. It may well be that even relatively new kit will already be out of warranty.
But the fact is, many organisations will feel their in-house staff, or their existing service partners, are entirely capable of upgrading memory and storage without causing a problem. Needless to say, it makes sense to discuss your plans with a specialist memory and SSD supplier who can help you select the best components for your workflows and workloads.
And your upgrade plan will still mean a conversation with your friends in finance. But they are likely to find the prospect of a super deduction on a third-party storage and memory upgrade far more palatable than okaying a riskier and undoubtedly pricier rip and replace program for the company’s entire fleet of servers.
Sponsored by Micron