W32‐1 meeting notes

Due date: 17 Sep 2024 (W38-2)

Today: 1 Aug (W31-4)

• General writing quality revisions

  • Better articutale what exactly is our novelty and contribution

• Novelty/claims

  • Cache sharing between volumes - both in terms of storage pooling and shared objects
  • Decoupling storage performance from capacity
  • Independent elasticity in both performance and capacity
  • QOS?
  • Tiered storage is a viable architecture - nvme for performance, hdds for capacity
  • Reduced networking demand
  • Separate out the endurance requirement for write logs

• Run experiments on new hardware (not final, just list of what we might want to run)

  • random read/write on small/large cache
  • ranod read/write scaled to multiple clients
  • VM boot times with and without optimisation
  • really sketchy numbers about backend cpu utilisation
  • bare-metal?
  • alibaba trace replay

• These numbers are comparing:

  • LSVD
  • RBD
  • LSVD, mounted with the RBD sharedobject interface
  • Tentative: Longhorn mounted over iSCSI
  • Tentative: Mayastore over NVMF

Plan for Alibaba trace replay:

• Select a target nominal total disk sizes at roughly 10-20TB
• Target IOPS at 100-200k
• Probable architecture: one fio process per virtual disk

Actions required:

• Isaac
  • Set up RPC for LSVD
  • Clean up the list of claims
  • Determine concretely the targets of our comparisons
  • From there, determine
• Kristi/timo
  • Setup test cluster
  • Setup nvmf gateway + fio clients
• Sumatra
  • Convert block traces into fio format
  • Figure out access to the nodes