Home · Servers for database applications
Workload guide
Servers for database applications
Databases buy three things from hardware: fast cores for query work, memory for the buffer pool, and storage latency lower than everything else. This guide matches those three to the server range we supply.
Follow the latency
A database server's storage path decides its real speed. The configurations in our range that serve databases hardest are the NVMe-heavy ones — the PowerEdge R960 listing, with 32 front NVMe bays on a four-socket chassis, and the DL380 Gen11's second listing with up to 36 NVMe drives — because they put the hot tables on the fastest media the chassis offers, under the largest memory the line carries.
Four-socket machines like the R960 earn their place where a single instance must scale up: very large buffer pools and per-core-licensed engines that reward fewer, faster sockets. Most database estates scale out on 2U dual-socket hosts instead, and the mainstream of the rack range is built for exactly that.
Memory is the cheapest performance
Every gigabyte of buffer pool is a disk read that never happens. Database briefs should state the database size and the hot fraction; the host's memory configuration is then sized to hold the hot fraction with room to grow. It is the single most reliable sizing rule in this guide, and it is why database hosts are quoted memory-first.
What to send
- The engine and the database size, plus the hot working set.
- Whether the design scales up (one big instance) or out (many hosts).
- The storage latency the workload needs — NVMe local, or shared array.
Backup and recovery are hardware decisions too
The database conversation usually ends at query performance, but the restore is the moment the hardware earns or loses its reputation. Backup throughput is a storage and network property sized at purchase; recovery time is an architecture property — how much data must move, onto what, at what speed. Both belong in the database brief alongside the working set: the data volume, the backup window, and the recovery time the business actually requires.
Estates that state those numbers buy coherent systems: hosts, storage tiers and backup targets sized as one design. Estates that omit them discover the gap during the first real restore, which is the worst classroom in computing. The question costs one line in the inquiry.
Licensing, once more, with feeling
Database platforms carry the sharpest version of the per-core licensing question, because the engines that dominate this workload are commonly licensed by the core. A four-socket host and a pair of two-socket hosts can run the same database at the same speed with licence exposures that differ by a factor of two. The hardware quote cannot settle that arithmetic — but the brief that names the engine lets the configuration be shaped with it in view, and the estate's software owner belongs in the sizing conversation before the configuration freezes, not after the invoice for the licences arrives.
Questions buyers ask
Is a four-socket server worth it?
Where per-core licensing or a single huge instance drives the design, yes. Elsewhere two 2U hosts usually beat one 4-socket machine on value.
Local NVMe or a storage array?
Local NVMe for the hottest tier, arrays for shared and protected tiers. The storage guides cover the array side of the design.
Price a configuration like this
Send the model name, your target configuration and the delivery destination. We reply with a configuration-matched quote and the checks to run before shipment.