Columnar Cache Tiering to Cut IMU Reconstruction Overhead
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional database systems inefficiently manage in-memory units (IMUs) by either maintaining them only in volatile memory or deleting them entirely when evicted, leading to inefficient data access due to the all-or-nothing approach, especially when multiple tiers of storage with varying performance characteristics are involved.
Innovation Solution
Implement a method to migrate IMUs across different tiers of storage (DRAM, NVRAM, NVMe FLASH, and disk) based on access frequency, maintaining mirror-format copies across lower tiers to reduce reconstruction overhead and increase storage availability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If IMUs are maintained only in volatile memory, then data access speed is improved, but storage capacity and cost are worsened
Solution Approach 1:
The patent segments the storage system into multiple tiers (volatile memory, NVRAM, NVMe FLASH, disk) with different performance characteristics. IMUs are distributed across these tiers based on access frequency, with hot IMUs in faster storage and cold IMUs in slower storage, resolving the contradiction between speed and storage capacity.
Solution Approach 2:
The patent adds a new dimension to storage management by introducing intermediate storage tiers (NVRAM, NVMe FLASH) between volatile memory and disk. This creates a multi-dimensional storage hierarchy that allows IMUs to be placed at different levels based on access patterns, achieving both fast access for frequently used IMUs and high capacity for less frequently used IMUs.
2Productivity
If IMUs are deleted entirely when evicted from volatile memory, then storage space is freed for new IMUs, but data access efficiency is worsened due to reconstruction overhead
Solution Approach 1:
The patent performs preliminary action by migrating evicted IMUs to lower storage tiers instead of deleting them. This preserves the IMU data in advance, so when the same IMU is needed again, it can be quickly restored from the lower tier without requiring full reconstruction, thus reducing reconstruction time while maintaining storage utilization.
Solution Approach 2:
The patent implements discarding and recovering by moving IMUs to lower storage tiers when evicted from volatile memory (discarding from high-speed storage) and then recovering them when needed again by loading back into the appropriate tier. This cycle allows storage space to be freed while preserving the ability to quickly recover the IMU data, avoiding the time cost of full reconstruction.
3Productivity
If mirror-format data is stored only in volatile memory, then query processing performance is improved, but storage cost and complexity are worsened
Solution Approach 1:
The patent segments the mirror-format data storage across multiple tiers based on IMU access frequency. Frequently accessed IMUs are maintained in volatile memory for optimal query performance, while less frequently accessed IMUs are stored in lower-tier storage. This segmentation achieves good query processing performance for hot data while avoiding the complexity and cost of maintaining all IMUs in volatile memory.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method comprises storing, within a storage system that is operatively coupled to a server machine that is executing a database server instance that manages a database, a persistent-format copy of a set of data, in a persistent-format, on persistent storage; wherein the set of data is from a table in the database managed by the database server instance. The method further comprises, in response to a load-triggering event, determining a target device in which to construct a mirror-format copy of the set of data, in a mirror-format; wherein the target device is one of the storage system and the server machine; wherein the mirror-format is different than the persistent format; wherein the target device is determined based on (i) metadata associated with the set of data, or (ii) one or more factors. The method further comprises, using data obtained from the persistent-format copy, constructing the mirror-format copy in volatile memory of the target device. One or more computer-readable media and a system comprise corresponding features.