With the announcement of vSAN 7 U1, VMware introduces several enhancements that improve the performance of vSAN and VMware Cloud Foundation (VCF) environments powered by vSAN. How much of an improvement? This will vary depending on the configuration and workloads, but VMware estimates customers may see up to 30% improvement in performance when comparing vSAN 7 U1 to vSAN 6.7 U3. Let's look at how this is achieved.
Doing more with what you already have
Efficiency and performance are often perceived and discussed as two different topics, but for the hypervisor and the services it offers, they are closely related. Improvements in efficiency assists with performance gains in two important ways.
- Minimizes impediments for I/O traversing the storage stack. This allows the operation to be completed faster, improving performance. Ensuring an optimal data path and management of metadata is key to delivering top-line performance.
- Minimizes resource usage. Keeping the usage of CPU, memory, and I/O amplification to a minimum allows for more VMs to run in an environment. Without well-designed software, inefficiency here could cause an excess of resources used on the host to the degree that there would be little left for the VMs it serves.
This helps explain why among other reasons, VMware chose to integrate vSAN into the vSphere: It provides the optimal performance, efficiency, and integration into the industry's premiere hypervisor. VMware always seeks to add new capabilities while optimizing the stack. It is of such high importance that our internal performance testing tracks the CPU cycles per I/O, which is a primary indicator of efficiency. Keeping this to a minimum allows for maximum VM-per host density and optimal performance.
Better performance through improved efficiency in vSAN 7 U1
Performance improvements rarely come from a single reason. vSAN 7 U1, along with a few supporting changes made in vSAN 7, incorporates enhancements across various aspects of the hypervisor that not only help with the hardware of today but the technologies of tomorrow. It is not uncommon for higher-performing hardware devices to shift bottlenecks, and expose opportunities for VMware to optimize the software, helping our customers realize the full benefits of the latest hardware. Let’s take a look at where these optimizations were made.
This feature was introduced to accommodate workloads that were not ideally suited for deduplication techniques. Offering a space efficiency feature that uses just compression allows for faster destaging of data from the buffer tier to the capacity tier. This increases the effective steady-state performance of a cluster in comparison to the same cluster running deduplication and compression. Steady-state performance not only helps I/O intensive workloads with large working-sets and long duty cycles but also large resynchronization events such as host evacuations.
Deduplication and Compression optimizations
In vSAN 7 U1, the deduplication engine was offloaded to another CPU thread, which in combination with other destager and component manager scaling enhancements improves the effective steady-state performance (provided by the capacity tier) of a cluster running deduplication and compression. This will benefit configurations that see contention during the destaging process, driving improved steady-state performance, and consistency.
Component manager scaling
In vSAN 7 U1, improvements have been made to the layer of vSAN responsible for managing and scheduling some aspects of access to the chunks of data (known as "components") that comprise an object. The number of component managers has increased, which helps satisfy the demand of higher IOPS and throughput to ensure that I/O will not be bound by that layer of vSAN as customers introduce newer, faster storage devices to their environment.
Durable writes during maintenance mode events
While the intention of this enhancement focuses on improving the durability of changed data while a host is in maintenance mode, it offers the potential to dramatically reduce the amount of data needed to be resynchronized after the host exits out of maintenance mode. Pair this with the reduced computational effort in merging the updated data versus object component rebuilds, and the effective reduction of resynchronizations can be significant. This will result in fewer resources used during these types of events, and the potential for more consistent levels of performance.
Multiple receive threads in VMkernel NIC (vmknic)
This enhancement increases the number of CPU threads available for use by the vmknic on the host that is responsible for vSAN network traffic. Assuming the network is not bound by the physical limitations of the NICs or switchgear, this will drive improved performance for front-end traffic such as VM workloads, and back-end traffic such as resynchronizations. The improvements from this can be most beneficial to workloads using RAID-5/6 erasure coding, mirroring using an FTT=2 or FTT=3, workloads using smaller I/O sizes, and hosts using multiple disk groups.
Fast host restarts
For fast and efficient I/O processing, vSAN uses an in-memory metadata table to track blocks of data as they are being committed to the caching tier. vSAN 7 U1 uses a new technique for graceful host restarts that commits this table to the host's caching device so that it can be quickly recreated to its last known state. For planned host restarts, this method offers significantly reduced time it takes to prepare the host for processing I/O when compared to the previous methods used.
Quantifying the level of improvement
The level of improvement realized in an environment will depend on the configuration, hardware, topology, and the workloads running on a vSAN cluster, as these factors influence the amount of improvement each discrete enhancement provides. For example, the vmknic optimizations may not be fully realized if the physical network is already saturated due to running at lower network speeds. Multiple component managers may see more benefit when faster storage devices are used and are under heavy demand. The hypervisor can only power workloads as fast as the underlying hardware allows, but thanks to the architecture it is easy to make incremental improvements through new hardware in the servers.
While many of the improvements described above provide substantial improvements for an element such as throughput for large sequential and medium-sized random writes and lower latency for guest VMs, performance gains are also achieved through other means. For example, the enhanced durability of writes may significantly reduce data movement for resynchronization activities. Fewer resources are used, which may provide improved levels of consistency in performance for the workloads in your cluster.
As the capabilities of hardware technology advance, so does the potential of greater performance of the software that allows for the hardware to be used in useful ways. VMware is excited about these enhancements in vSAN 7 U1. They help you do more with what you already have while being ready for the latest in hardware technology. The result is better performance through improved efficiency.