High-memory servers
Some workloads are bounded by memory, not by cores. When the dataset has to sit in RAM, the amount and the layout of that RAM is the whole decision.
What actually needs it
In-memory databases, large caches, analytics that hold a working set resident, and build farms that fan out across many parallel jobs. For these, spilling to disk is the failure case you are paying to avoid.
Size for the working set plus headroom
Measure the resident set under real load, then leave room for the page cache, the operating system and growth. A box that runs at 98 percent memory has no margin for a query that is heavier than usual — it reaches for swap or the out-of-memory killer, and both are worse than a bigger spec.
ECC and NUMA
Every machine we supply at this tier uses ECC memory, which corrects the single-bit errors that are statistically certain at hundreds of gigabytes — see memory errors. On dual-socket boxes the RAM is split across sockets, so pinning matters: NUMA and huge pages covers how to keep a process close to its own memory.
Still stuck? Mail support@novogara.com — an engineer answers, at any hour. Back to the knowledge base