About this article: This article is reviewed against the cited sources. This is source research, not lab validation. Check the stated versions and your own requirements before making changes.
A server is oversized when its busiest periods across a full business cycle still leave a large share of its processors and memory unused. To reduce it safely, collect peak figures over that whole cycle, including month-end and the end of the financial year, take away a modest step rather than the whole surplus, book the change with a planned restart, watch the next peak, and keep the old size written down so it can be put back quickly. On a busy VMware host, removing processors a server cannot use can also help its performance, not just its cost.
Why it matters
Where a hosting plan charges by allocated processors and memory, an oversized server costs money every month. Servers can grow over the years through "just in case" requests that nobody revisits. But reducing a server on the strength of a quiet week is how a business discovers, at end of quarter, why it was that size. The method matters more than the saving.
How it works
Evidence over a full cycle
Cloud tools that recommend smaller sizes make the evidence period explicit. Microsoft's Azure Advisor notes that "The available lookback periods are 7, 14, 21, 30, 60, and 90 days". AWS Compute Optimizer similarly lets you "Configure a specific lookback period option" and "Adjust both the headroom and threshold of your CPU utilization".
Broadcom's own procedure asks for baseline data "over a representative 24-hour period". Our recommendation is deliberately stricter, because a day, or even 90 days, can miss the yearly peak of a business server: before reducing anything, check that your evidence covers the busiest events the server faces, such as payroll, BAS, end of financial year or the season. If you have only a quiet period, wait. Our published article Sizing a business application server from measurements, not the vendor's minimum covers collecting figures across a full business cycle.
What oversized looks like
Broadcom's rightsizing guidance for ESXi 8.0 gives the signals:
- Memory: "A consistently low active-to-allocated ratio suggests oversizing."
- Processors: for servers with several virtual processors, co-stop is a waiting measure. Broadcom says co-stop "Below about 3% per vCPU is normal", and that sustained higher values "typically indicate the VM has more vCPUs than it can effectively use, and reducing the vCPU count usually helps". For processor use inside the server, the same article says to "Aim at or below 80% during normal operation, with headroom for spikes".
Two other figures point the other way, at the host rather than the server. Broadcom notes that "Sustained high ready time generally signals CPU contention at the host level rather than undersizing of the individual VM", and that "Non-zero ballooning indicates host memory pressure has caused ESXi to reclaim memory from the guest". If the server is slow and these are high, look at the host as well as the server: the provider must address a busy host, and shrinking one server will not cure that on its own, although rightsizing oversized servers can reduce the total demand on the host. Diagnose both rather than assume either.
Your provider can read these platform figures; ask for them alongside the Windows peaks. Our published article CPU overcommit in hosted clouds: ready time and NUMA explains ready time in more depth.
When fewer processors can help
It sounds wrong, but Broadcom advises: "Configure neither more vCPUs than the workload can effectively use, nor fewer than it needs." It explains that "extra vCPUs add scheduling overhead and can reduce performance on heavily loaded hosts". The gain, where there is one, comes on heavily loaded hosts and shows as high co-stop; on a lightly loaded host there may be none.
What to choose and why
A safe reduction method
- Record the current size and the reason it was set, if known.
- Collect a full cycle of peaks: sustained processor peak, memory in use, and the platform figures above.
- Check what is special. A database server's memory figure can look full while it is working normally (see below), so it is not evidence either way. Ask the application owner about licence terms tied to processor counts.
- Cut one step. For example, from eight processors to six, not to four; from 32 GB to 24 GB, not to 16 GB.
- Book a planned restart. Broadcom's procedure says "Changes to vCPU count, memory size, or CPU topology require the VM to be powered down".
- Watch the next peak. If users notice slowness, restore the recorded size; that is also a shutdown change, so have a window ready.
- Repeat only if the next cycle again shows a clear surplus.
Illustrative example
A professional-services firm's remote desktop server has 16 processors and 64 GB. Over a year including June, its sustained processor peak is 35 per cent and memory in use peaks at 30 GB, with low active memory on the platform charts. The firm reduces to 12 processors and 48 GB in a Friday evening window, watches the next month-end and BAS quarter, sees peaks of 45 per cent and 31 GB with no complaints, and stops there. The example is illustrative.
Read next
Related articles you may want next.
- HostingSizing a business application server from measurements, not the vendor's minimumWhich Windows performance counters to collect on an existing server, for how long and at what interval, and how to turn the results into processor, memory and disk sizes for a hosted replacement.Read article
- PerformanceCPU overcommit in hosted clouds: ready time, NUMA and how to size honestlyMeasure CPU wait on ESXi, Proxmox VE and Hyper-V, read ready time correctly, keep VMs inside NUMA boundaries and size vCPUs from evidence.Read article
- HostingSQL Server licensing in virtual machines: per-core minimums, Software Assurance and hosting providersHow Microsoft counts SQL Server core licences for a virtual machine, why a small VM still needs four, what changes when a hosting provider supplies the licence or you bring your own, and when a standby server needs licensing. General information, not legal advice.Read article
Risks and limits
- Short evidence misleads. A week or a month without a peak proves nothing about the peak.
- Memory is not one number. SQL Server "can change its memory requirements dynamically based on available system resources" and, under normal circumstances, will "claim memory up to the limit set by max server memory (MB)"; read in-use and active figures, not just "used".
- The guidance is for one platform. Broadcom's figures apply to ESXi 8.0; Hyper-V, Nutanix, KVM and other platforms have their own guidance and figures.
- Licences can depend on size. Some software, such as SQL Server, can be licensed per processor core; reducing may change what you owe, but check the terms before and after. Our article SQL Server licensing in virtual machines: per-core minimums, Software Assurance and hosting providers explains per-core minimums.
- Restarts are part of it. A reduction is a planned interruption; do it in a window and have the old size ready to restore.
- Cloud tools are a method, not a verdict. Recommendations from Azure or AWS tools apply to their platforms; the principle of evidence periods and headroom transfers, the numbers do not.
Sources
Opened 15 November 2026; Broadcom 438023 and the SQL Server memory page reopened 16 November 2026.
- Broadcom Knowledge: Rightsizing virtual machines on ESXi 8.0: vCPU, memory, and CPU topology guidance, article 438023.
- Microsoft Learn: Optimize virtual machine (VM) or virtual machine scale set (VMSS) spend by resizing or shutting down underutilized instances, updated 1 April 2026.
- Microsoft Learn: Server memory configuration options, page shows last updated 11 May 2026 (page metadata: modified 19 October 2026).
- AWS: Rightsizing recommendation preferences, AWS Compute Optimizer User Guide.
Related articles: CPU overcommit in hosted clouds: ready time and NUMA; Sizing a business application server from measurements, not the vendor's minimum.
Related service: Managed support.
Learning notes
- Collect peaks across a full business cycle before reducing anything.
- Low active-to-allocated memory suggests oversizing; co-stop above about 3% per vCPU suggests too many processors.
- High ready time or ballooning points at a busy host; diagnose host and server together, since shrinking one server alone will not cure host contention.
- Extra virtual processors can slow a server on a busy host.
- Reduce one step at a time, in a planned window, and watch the next peak.
- Write down the old size so it can be restored quickly.
This article was originally published on .