Persistent Memory Accelerators
Overview
Exadata X8M introduced persistent memory (PMEM) data and commit accelerators in front of flash cache, enabling orders of magnitude lower latency accessing remotely stored data.
Persistent memory is a modern silicon technology, adding a distinct storage tier of performance, capacity, and price between DRAM and Flash. As the persistent memory is physically present on the memory bus of the storage server, reads perform at memory speed, much faster than flash. Writes are persistent, surviving power cycles, unlike DRAM.
Persistent Memory Data Accelerator
Database uses RDMA instead of IO to read remote Persistent Memory. By utilizing RDMA to access persistent memory remotely, Exadata Persistent Memory Data Accelerator bypasses the network and I/O stack, eliminating expensive CPU interrupts and context switches, reducing latency by 10x, from 200 microseconds to less than 19 microseconds for 8K database reads.
Persistent Memory Commit Accelerator
Log Write latency is critical for OLTP performance. A faster log write means faster commit times. Any log write slowdown stalls the whole database. Persistent Memory Commit Accelerator allows the database to issue a one-way RDMA write to PMEM on multiple Storage Servers. This bypasses network and IO software, interrupts, context switches, resulting in up to 8x faster log writes.
Persistent Memory in Shared Storage Advantages
Adding persistent memory to the storage tier means the aggregate performance of this new cache tier can be dynamically used by any database on any server. This is a significant advantage over general-purpose storage architectures, which preclude sharing storage resources across database instances.
Smart Exadata System Software also ensures data is mirrored across storage servers, which provides additional fault-tolerance. Exadata’s unique end-to-end integration between Oracle Database and Exadata Storage Servers automatically caches the hottest data blocks efficiently between the database buffer cache, persistent memory, and flash cache.
- Better utilization of PMEM capacity – only hottest data moved into PMEM
- More adaptive - aggregate performance of PMEM on all storage servers dynamically used by any server
- End to end security – PMEM only accessible to databases using database access controls, no OS or local access
- Simpler – all PMEM benefits with no extra management
- More resilient – automatic mirroring of PMEM across storage servers protects from PMEM failures
- Much better latency – RDMA from Database directly to PMEM