Caching VMware storage IO to in-host Intel Optane SSD and host RAM improves VDI latencies

Use Case:
VDI Performance

Location:
Daejon, Korea

Challenges:

  • Their CEPH storage software was exhibiting high write latencies from time to time.
  • VDI VMs being end user facing, their IT department needed consistently low read and write latencies in every VM and at all times. They couldn't afford even one blip over 10ms latency, else that would result in support calls.

Benefits:

  • By caching all VM reads and writes to ESXi host RAM / Optane NVME, VirtuCache ensures that storage latencies at the VM level are less than 1ms for every VDI VM, and at all times.

Caching VMware storage IO to in-host Intel Optane SSD and host RAM improves VDI latencies

The Virtunet Difference
The customer selected VirtuCache because:
  • VirtuCache is the only solution in the market that improves storage performance without requiring a wholesale replacement of customer's storage infrastructure.
  • The only other alternative was to replace CEPH storage with a faster All-Flash array like the EMC XtremIO, which would have been very expensive.
  • February 3,2021; Daejon, Korea.

    Chungnam National University Hospital (CNUH) uses VirtuCache with Intel Optane SSD and host RAM to improve VMware Horizon VDI performance.

    CNUH VDI Environment.

    Chungnam National University Hospital (CNUH) is a 1400-bed full-service hospital located in Daejeon, Korea. CNUH’s VDI environment uses VMware’s Horizon software to serve virtual desktops to 600 end users. Horizon is running in two ESXi clusters that has 19 hosts. Open source CEPH storage software running in VMs is used to pool together locally attached SATA SSDs in ESXi hosts and present this storage over iSCSI to the ESXi hosts.

    High Storage Latencies at CNUH.

    It is common knowledge that CEPH storage software exhibits high latencies, especially for writes. This is because of high software overhead in CEPH, which in turn is due to CEPH replicating all the writes to one or more nodes (for data protection). If CEPH is run in VMs, as is the case at CNUH, you have the additional overhead of VMware’s VMFS filesystem. Even though CNUH was using enterprise SSDs with CEPH, the write latencies were high from time to time.

    VirtuCache configuration to reduce VDI VM latencies.

    VirtuCache automatically caches all data from VMware VMs to any in-host SSD or RAM. At CNUH, VirtuCache is configured to cache to 64GB RAM per host in one ESXi cluster and to 375GB P4800 Intel Optane SSD per host in the other ESXi cluster.

    In VirtuCache, you can apply three caching policies.

      ‘Write Through’ Caching policy that caches reads but not writes.

      ‘Write-Back with One Replica’ caching policy that caches reads and writes, and it mirrors (replicates) writes to one additional host in the ESXi cluster. VirtuCache does this to protect against data loss if the host fails.

      ‘Write-Back with No Replica’ works like the ‘Write-Back with One Replica’ caching policy, except that it does not replicate writes to another host, which means that if a host fails then there will be data loss for the VMs that have this policy.

    The customer has deployed non-persistent VDI VMs using Horizon Linked Clones. Since these VDI VMs are non-persistent, data loss for these VMs could be tolerated in case of host failure. Hence the ‘Write-Back with No Replica’ caching policy is applied to these VMs. The advantage to this policy versus the ‘Write-Back with One Replica’ policy is that there is no network overhead involved when it comes to caching writes since the writes are not replicated. As a result the write latency is equal to the latency of the cache media, versus in the case of ‘Write-Back with One Replica’ policy, where the VM write latency is the sum of network latency and cache media latency.

    Note that if you use ‘Write Back No Replica’ policy with host RAM, it results in the lowest latency VDI VMs possible, since RAM is the lowest latency media (even lower than NVME SSDs) and there is no network overhead associated with write latencies.

    For VDI management VMs – Connection Server, Composer Server, and vCenter VMs, ‘Write-Back with One Replica’ caching policy is used, since data loss cannot be tolerated for these VMs.

    Performance Improvement with VirtuCache.

    Cache Media Type.

    VM Latency Before VirtuCache.

    VM Latency After VirtuCache.

    Host RAM as Cache

    20ms

    0.1 ms. Here VirtuCache is caching to RAM

    Intel’s Optane SSD as Cache

    20ms

    1.0 ms. Here VirtuCache is caching to Optane NVME SSD

    Download Trial Contact Us