Skip to content

Cold Storage.md

Latest commit

 

History

History
53 lines (51 loc) · 2.66 KB

Cold Storage.md

File metadata and controls

53 lines (51 loc) · 2.66 KB
Raw

Definition

Systems for storing large volumes in infrequently accessed data.

  • Include archival data (e.g. data stored for regulatory purposes that is rarely accessed for business needs)
  • Needs very high capacity, cheaply.
  • Cost is important - the data is infrequently accessed, so unlikely to contribute nto company profit
  • Access can be very slow For a Write Once - Read Occasionally/WORO workload.

Provisioning

Often using very cheap disks (e.g. [[Shingled Magnetic Recording]]), or magnetic tape, or even blu-ray disk. We only need the minimum resources required for the write & rare read workload.

  • Only need power, cooling required for the small read workload (not for massive access)
  • Bandwidth can also be small (we do not care about slow access)
  • Only need disks to be on during access.

Architecture

A CSD rack with groups of disks, at any time one group can be spun up.

  • switching group on order of $10\to30$ seconds.
  • Need to track placement of data
  • Random access is very slow

Examples

Microsoft Pelican

A 52U rack with 1152 archival class $3.5$ inch SATA disks each with avg $4.55$ TB storage for total $5$ PBs of storage.

  • Only 8% of disks are active at a given time. Systems on board (domains) supply resources (power, cooling, bandwidth) to a subset of disks.
Domain-Conflicting Disks that are in the same resource domain
Domain Disjoint Disks with no shared resources
  • Disjoin can be simultaneously active, conflicting potentially cannot.
  • Disks are split into groups (all disks in a group can be concurrently active / are disjoint), each blob/file is split into $k$ $128$KB fragments
  • I/O scheduler can reorder operations to amortize the spinup latency of groups.

Advantages

  • reduced costs & power consumption

Disadvantages

  • Tight constraints
  • No thorough justification published by Microsoft / maybe not optimal?
  • Sensitive to hardware changes

Others

  • OpenVault's Knox
  • Facebook Cold storage
  • Amazon Glacier

Data Processing

Load

When querying data, software needs to batch requests and cache to reduce the number of expensive group switches.

Store

Data must be flushed back to the [[Cold Storage]] once the buffer is full, while reducing the number of group switches.

Method Description
Buff-Pack (Greedy Algorithm) Flush to current disk group, then switch to disk group with most free capacity
Off-Pack Flush the entire buffer to disk, can write to any disk group, then transfer later.
Off-Pack is better for smaller buffers, higher number of disk groups and for higher throughput.