If hardware price squeezes make you sad, VMware says it will all end in tiers

VMware is working to find new roles for its memory tiering technology, in the expectation that memory and hardware prices will remain high for years and users will therefore look for ways to put off server purchases. Memory tiering moves data out of DRAM and onto another storage medium that’s fast enough that CPUs can still access info at speeds that don’t slow application performance. VMware’s implementation works by looking for “cold pages” in DRAM – chunks of memory containing data that a CPU hasn’t touched and/or won’t need for a while. When a CPU can’t find any capacity in DRAM, it can pause operations until memory becomes available. That leaves CPUs and expensive DRAM idling wastefully when they could be churning through work. To stop the waste, memory tiering looks for cold pages in DRAM and moves them to another storage medium – for VMware that means fast and resilient NVMe drives. The virtualization pioneer introduced memory tiering last year in version 9 of Cloud Foundation, its bundle of products that allow users to assemble and operate a private cloud. In this year’s VCF 9.1 update, VMware tweaked it to improve performance when running databases – always a key workload – and added dashboards that show users the extent of tiering activity, plus more management tools. VMware delivered memory tiering just as the AI boom saw the price of DRAM and solid state storage skyrocket. DRAM now often costs several times more than the servers it lives in. At the VMware Explore conference on Monday, staff technical marketing architect Dave Morera said VMware plans to support more VM types with the tech. “We have a two-to-three year roadmap,” he said in a conference session. Some of the items on that roadmap will emerge in the VCF 9.2 release that VMware will likely debut around May 2027. Morera said VMware’s memory tiering currently supports about 75 percent of workloads and wants to push that number higher by supporting Microsoft’s Radius cloud-native application platform and so-called “monster VMs” that VMware allows to run up to 960 vCPUs and 16TB of memory across several physical hosts. One of VMware’s crown jewels is vMotion, the tech that moves VMs from one cluster to another with almost no downtime. Memory tiering currently supports vMotion, but Morera said he wants to improve performance by solving some unidentified physics problems. Memory tiering is built into VMware’s ESX hypervisor. Using it requires NVMe drives capable of at least 100,000 writes per second, and 7,300 TB of writes across their lifetime. Another metric commonly applied to NVMe is Drive Writes Per Day (DWPD) – the number of times it is possible to completely overwrite a drive in a single day, without the device dying before its warranty expires. VMware recommends drives with a DWPD of at least three. It’s possible to use memory tiering in a 1:4 ratio – 1 GB of DRAM and 4 GB of NVMe. VMware recommends a 1:1 ratio and claims that it can lead to applications consuming 30 percent fewer CPU cycles, and 40 percent lower cost of ownership by allowing users to “densify” their infrastructure into fewer servers. The Broadcom business unit also argues that adopting memory tiering can improve a server’s performance and therefore extend its working life, making it possible to defer hardware purchases. Company execs argue that’s a win given the long lead times for servers, the high cost of DRAM, and the likelihood that server vendors will need to change quoted prices due to market fluctuations. Of course VMware would say that, because it wants more people to implement VCF. Yet it seems VMware customers and prospects are very interested in memory tiering, as the two sessions on the topic that The Register to gather info for this story filled 500-seat conference rooms. ®