Access Official Technical PDF

Share Summary:LinkedInEmail

SSD caching in virtualization can help reduce the performance gap between high-capacity hard disk drives (HDDs) and faster solid-state drives (SSDs). But adding an SSD cache is not enough on its own: cache space must be shared effectively among virtual machines (VMs), and the way writes are handled can affect both performance and SSD endurance. ECI-Cache is a hypervisor-based approach designed to account for these factors together.

Rather than assigning cache space based only on how frequently data is reused, ECI-Cache considers whether an access is a read or a write. It also selects a write policy for each VM according to its workload. The result is a way to manage a shared SSD cache with attention to performance, cost efficiency, and the number of writes sent to the SSD.

Why virtualized storage needs workload-aware caching

HDDs offer capacity at relatively low cost, but their performance can be limited for random I/O. SSDs can serve storage requests more quickly, yet using them to replace all HDD capacity can be expensive. An SSD cache placed in front of HDD-based storage offers a compromise: frequently or usefully reused data may be served from the faster tier, while the HDD subsystem continues to provide bulk storage.

In a virtualized server, many VMs can generate different I/O patterns while sharing the same storage resources. A single global cache can let one VM consume cache space that would otherwise help another. Static partitions avoid that competition but may reserve space for workloads that do not use it effectively, and do not readily adapt when VMs are added at runtime. Dynamic partitioning is more flexible, but cache sizing alone does not address the effects of write behavior.

Those effects matter because cache policies make different trade-offs. Write-back (WB) stores writes in the cache and later propagates them to the storage subsystem. It can improve read and write performance, but a failure before buffered data is propagated can result in data loss. Write-through (WT) sends writes to both cache and storage, providing a different reliability trade-off, while read-only (RO) caching sends writes directly to storage and caches read data. RO does not accelerate writes, but it avoids buffering them in the SSD cache.

Useful Reuse Distance: sizing cache for valuable reuse

Traditional reuse distance estimates how far apart accesses to the same address occur. Cache sizing based on this measure can treat every access as evidence that a block may be worth keeping. However, some write sequences do not lead to a later read from the cached block. Storing those writes can consume cache space and add SSD writes without improving the read hit ratio.

ECI-Cache introduces Useful Reuse Distance (URD), which accounts for request type. It focuses on access sequences that can contribute to read hits:

  • Read after read (RAR): A later read may find data fetched into the cache by an earlier read.
  • Read after write (RAW): A later read may use data previously written to the cache.

By contrast, write-after-read (WAR) and write-after-write (WAW) sequences do not necessarily produce a later read hit. URD excludes these patterns from its reuse-distance calculation. In the paper’s illustrative access sequence, traditional reuse distance indicates a cache allocation of five blocks, while URD indicates two blocks while preserving the example’s hit ratio. This example explains the principle; actual allocations depend on each workload’s observed access pattern.

Smaller allocations can make cache capacity available to other VMs and reduce writes associated with retaining blocks that provide no useful read reuse. When aggregate estimated demand exceeds SSD capacity, ECI-Cache uses hit-ratio estimates to allocate the available space with the aim of minimizing aggregate VM access latency.

Per-VM write policies reduce unnecessary SSD writes

ECI-Cache pairs cache sizing with workload-specific write-policy selection. Its hypervisor-based design monitors I/O requests and associates them with the VM that issued them. An analyzer uses the observed request types and access patterns to guide cache-size and policy decisions; an actuator applies those decisions and routes requests through the cache or to the storage subsystem.

The approach distinguishes writes likely to be read later from writes that appear to have no further read reuse. It can use WB where buffered writes are useful to subsequent reads, and RO where avoiding write caching is preferable. The policy is reassessed periodically as workload behavior is monitored. This is important in consolidated environments, where applying one policy to every VM can expose some workloads to unnecessary SSD writes or fail to provide a performance benefit.

Policy selection does not eliminate the need to consider reliability requirements. WB can improve performance, but buffered writes introduce a risk if the SSD fails before those writes reach the storage subsystem. RO avoids buffering writes in the cache, but does not provide write acceleration. The appropriate choice therefore depends on workload behavior and the system’s requirements.

What the reported evaluation shows

The researchers implemented ECI-Cache in QEMU, an open-source hypervisor, and evaluated it using more than 15 workloads from SNIA Microsoft Research traces. Their reported comparison with the state-of-the-art dynamic cache-partitioning scheme showed a 17% performance improvement, a 30% improvement in performance per cost, and a 65% reduction in writes committed to the SSD.

These are results from the study’s experimental setup, not a guarantee of the same outcomes in every deployment. The practical value of the approach is its combination of per-VM workload monitoring, request-aware cache sizing, and adaptive write-policy assignment. Together, these mechanisms address cache utilization and SSD write activity rather than treating cache capacity as the only management decision.

Practical implications for automation and infrastructure teams

For engineers and technical decision-makers managing virtualized infrastructure, ECI-Cache highlights several considerations when evaluating SSD caching:

  • Assess read and write behavior for each VM instead of assuming a shared policy will suit every workload.
  • Distinguish useful read reuse from writes that are unlikely to produce later cache hits.
  • Consider cache capacity, write policy, performance, cost, and endurance as connected decisions.
  • Review reliability requirements before using a policy that buffers writes in the SSD cache.

ECI-Cache’s central lesson is that effective SSD caching in virtualization depends on more than allocating space to workloads with frequent reuse. By accounting for request type and adjusting cache size and write policy per VM, operators can make cache management more responsive to workload differences. The reported results demonstrate the potential of this approach, while deployment decisions should still be validated against the I/O patterns and reliability needs of the target environment.

Access Official Technical PDF

About The Author