Upgrading server hardware vs buying the latest platform

#KingstonCognate introduces Simon Besteman

Simon Besteman is a French and Dutch national and a veteran of the ICT and data center industry. He has over 20 years of experience working at ISPs, suppliers, and data centers and as a management consultant for a wide range of organisations. Simon has held various senior management positions in multinational companies in the areas of service, sales and marketing, operations management, and strategic development.

Currently, Simon is the Managing Director of the Dutch Cloud Community, the Dutch coalition of hosting providers. As a leading representative of the industry, he is a frequent blogger on industry and policy matters, a keynote speaker at congresses and conferences, and a participant at Dutch government round tables on telecommunication, data center and internet regulatory matters. He sits on the boards of various industry groups, with a focus on education, employment, and governance.

Extending life through memory

Before upgrading, it is important to identify the performance bottlenecks in your current setup for your specific application. One potential bottleneck may be memory usage.

Note that fully populating the memory socket can affect the memory bandwidth. Memory is always a trade-off between speed and capacity. Think of it as a road that would have traffic made of trucks that carry a large volume of goods at low speed, and high-performance sports cars that have little space for luggage but can travel at high speeds.

The model of your CPU and the server platform, as well as the way memory modules are populated within the server, can affect the bandwidth. In most cases, higher memory bandwidth can be reached by only populating 1 DIMM per Channel (1DPC). Adding a second DIMM per channel often results in the memory clocking down. However, insufficient memory capacity in your system may have a higher impact on performance than lower memory bandwidth.

If your servers have free memory sockets, adding extra memory modules or upgrading to higher bandwidth memory modules (such as DDR4 3200MT/s) may improve performance. However, it is important to always follow the memory population guidelines provided by the system manufacturer for optimal performance and to reduce the risk of stability or compatibility issues.

Key considerations for upgrading your servers

Here are a few important considerations before upgrading your hardware:

• Check if your memory usage is getting full under normal workload. If this is the case and your servers still have free memory sockets, it will be worth considering adding extra memory modules (DIMM).
• Replacing existing memory modules with new ones that have higher bandwidth (e.g., replacing DDR4 2400MT/s with DDR4 3200MT/s) can also result in some performance improvement, depending on the application and assuming the CPU and host system support higher memory bandwidth.
• If you identify that the bottleneck comes from the storage rather than the memory, then upgrading the storage devices might be the right thing for you. If your current storage setup is only using the HDD storage medium, consider SSDs for better performance.

Storage Is Key to Extending the Lifespan of Your Servers

That said, replacing all HDDs may not be suitable for your application or may be too pricey. Therefore, using SATA SSDs for caching and HDDs for high-capacity cold storage may be enough (again, this will depend on your application).

Replacing HDDs with SATA SSDs is straightforward as they use the same interface and communication protocol (AHCI). But when considering using PCIe NVMe SSDs there are a few things to keep in mind:

• Verify that your motherboard and operating system support NVMe (most systems and OSs released after 2015 usually support NVMe out of the box).
• Identify the right SSD with the right connector and dimensions for your PCIe NVMe storage bay. Nowadays, PCIe NVMe SSDs in 2.5” form factor with U.2 connectors (such as Kingston DC3000ME) are quite common. Several server models from different vendors tend to support this form factor. Please note that 2.5” U.2 SSDs can have different thicknesses, either 7mm or 15mm. Make sure to get the right data center SSD that will fit in your drive’s bays.
• If your server does not have a drive bay that supports U.2 NVMe SSDs, you may be able to replace the existing drive bay that supports SATA/SAS drives with a drive bay for NVMe SSDs. In that case, the things that you would need to consider replacing are the drive cage (if the U.2 SSDs that you want or have do not fit in the existing one), the existing backplane with one that has a port/connection for U.2 SSDs, the RAID controller, and the cable connecting the backplane to the controller.

For even better QoS (IOPS consistency and low latency), it is better to consider PCIe NVMe SSDs as caching or for full storage pools. The performance difference between SATA AHCI and PCIe NVMe SSDs is impressive and can be critical for latency-sensitive workloads and applications such as AI, machine learning, online transaction processing (OLTP) databases, big data analytics, cloud computing, operational databases (ODB), database applications, and data warehousing. You can read more about NVMe from Kingston: The benefits of NVMe in Enterprise.