Rapid advances in server peripherals, such as networking and storage, create a dilemma for IT shops: should they replace or upgrade older servers? It pays to measure the alternatives, since this is the first time in recent industry history when sever upgrades can dramatically affect performance.
In outdated data centers with gigabit Ethernet and spinning drives, the processor spent a good deal of time in wait states. Drive access took roughly 10 milliseconds and networked storage couldn't keep up with demand. Most servers failed to reach their potential due to low I/O rates, causing server farms to be two to five times larger than they needed to be based on CPU capability alone.
Traditionally, IT teams could replace servers with the latest models and see performance improvements. The CPUs may have followed Moore's law, but the rest of the system didn't change. Disk drives still spun at the same speeds, and the network was still the data center's 1 GbE backbone.
Today, the availability of inexpensive solid-state drives (SSD) and the rapid evolution of Ethernet change the rules -- and should influence your server upgrade checklist.
Choosing the right server-storage model
Forklift replacements are expensive, so let's explore options for upgrading servers rather than replacing them. Storage is an obvious item to include on a server upgrade checklist, since it is often a major bottleneck. In server direct-attached storage, there is typically four to six enterprise hard drives. In aggregate, these deliver around 600 to 900 IOPS. Adding two inexpensive 1 TB Serial Attached Technology Attachment (SATA) SSDs to the storage pool can get you in the range of 135,000 IOPS -- more than enough performance to match most applications.
Most drives today have no wear-out issues with typical applications, though heavy load versions are available with higher write durability at a premium. Hard disk drive vendors now specify wear life for their spinning drives that is comparable with SSD specs, essentially eliminating the wear use issue. There is little value in selecting a dual-port Serial-Attached SCSI (SAS) SSD. In a typical server, the extra port only protects a cable -- the two ports share a connector and the SAS integrated circuit at the drive is shared, too -- but you'll pay six or more times the price for the drive.
SATA SSDs serve well enough as an upgrade, but ensure your system can support drives that are not branded by your server vendor. Dell, for example, sells both proprietary drives and less pricy drives from companies such as SanDisk or Micron. Using cheaper drives might partially void the server warranty, but that's unlikely to pose a barrier on older systems. There may be a special vendor driver that looks for proprietary drives to use, however, that does not recognize these third-party drives. This is more of an issue with SAS drives than SATA.
Network considerations between servers
The second item to add to your server upgrade checklist is the network. If your servers use the network for any serious traffic, 10 GbE is the low threshold today. For example, virtualized clusters can take serious bandwidth because of image loading. Container adoption will increase network load considerably. In many cases, with older systems, you'll need to add 10 GbE in the server, adding to the cost. If you are modernizing segments of your data center and installing a new LAN, consider 25 GbE instead of 10 GbE connections. Data center teams can expect the price of 25 GbE to go down in 2016 and beyond as adoption rises.
Faster LANs mean lower latency to Ethernet network storage. A few SSDs can work wonders in those storage boxes, but the tendency to have proprietary hooks in drive interfaces is stronger in storage. An alternative is an all-flash array, which can replace a few thousands of IOPS with several million. The fast LANS are necessary to make all-flash appliances reach their capability.
Final factors on a server upgrade checklist
The final thing to target when upgrading servers is memory. Memory demand largely depends on the workload. For example, database performance improves with in-memory mode. Oracle regularly claims 100x performance gains. Typically, a database requires at least 128 GB of memory in each server, so check if your server can hold that much. For other applications, once you tune the storage and LAN, a memory upgrade of 2x or 4x makes sense. For example, extra memory in a containers server could significantly increase the instance count.
The alternative to a server upgrade is a forklift replacement, which resets the timer on server life expectancy and introduces the newest mainstream technology. On the downside, it's expensive and disruptive. This is especially an issue as we head toward new server architectures in 2016 and 2017. We can expect faster memory, fully integrated nonvolatile dual in-line memory module and embedded 25 GbE ports, with SATA-express support. Getting two more years out of existing systems by upgrading servers could be the cheaper solution, especially if the SSDs are carried over to the new systems and the 25GbE LAN aligns with your strategic data center needs in one or two years.
Upgrading your server OS is a smart decision
Windows Server 2016 upgrade tip
How virtualization can help in a server upgrade