With the introduction of the X8M (and subsequently X9M and newer) generations, Oracle redefined the "speed of light" for database commits by introducing Persistent Memory (PMEM).

While earlier versions of Exadata relied on Smart Flash Logging (racing between Flash and Disk), the PMEM Log takes latency reduction to a molecular level by utilizing Intel Optane DC Persistent Memory and RoCE (RDMA over Converged Ethernet).

1. The Core Function: Ending the Log Write Stall

In a database, a "Commit" isn't a success until the Redo Log is safely written to non-volatile storage. Even with the fastest NVMe Flash, you still have to deal with the overhead of the storage software stack, interrupts, and context switches.

The PMEM Log acts as a dedicated, ultra-high-speed landing zone for these redo log writes. By using RDMA (Remote Direct Memory Access), the Database Server can write redo entries directly into the memory of the Storage Cell without involving the storage server’s CPU.

2. How it Works: The RDMA Shortcut

Traditional I/O requires a "handshake" between the database and the storage software. PMEM Log changes the workflow:

1. Direct Mapping: The database server knows exactly where the PMEM Log buffers are located in the remote storage cell's memory.

2. The Write: Instead of sending an I/O request, the database performs a Remote DMA Write. It pushes the redo data directly over the 100Gbps RoCE network into the PMEM.

3. Hardware Persistence: Because PMEM is non-volatile (persistent), the data is safe the moment it hits the memory module.

4. Instant Ack: There is no "wait" for the storage CPU to process the request. The database gets a hardware-level confirmation in under 19 microseconds.

3. PMEM Log vs. Smart Flash Logging

It is important to note that PMEM Log doesn't replace Smart Flash Logging; it supercharges it.

• In X8/X9M: The "race" is now between PMEM and Flash (instead of Flash and Disk).

• The Winner: PMEM wins 99.9% of the time. The Flash log remains as a fallback to ensure that if the PMEM area is momentarily full or busy, the commit latency still remains within an acceptable "Flash-speed" window.

4. Why 19 Microseconds Matters

To put this in perspective:

• Traditional SAN/Cloud: 1,000 to 5,000 microseconds (1–5ms).

• Exadata Flash: ~200 to 500 microseconds.

• Exadata PMEM Log: < 19 microseconds.

This 10x to 50x improvement in commit latency allows for massive scaling of OLTP applications. It removes the "Log File Sync" bottleneck that traditionally limits how many transactions a single database instance can handle per second.

[Image comparing latency of HDD vs Flash vs PMEM in Oracle Exadata]

5. Key Advantages

• CPU Savings: Since the Storage Cell CPU is bypassed during the RDMA write, those CPU cycles are freed up to perform more Smart Scans and complex queries.

• Predictability: PMEM doesn't suffer from the "garbage collection" or "wear leveling" spikes inherent in Flash technology.

• High Availability: Like all Exadata storage, PMEM logs are mirrored across multiple storage cells. A commit is only acknowledged once it is persistent in PMEM on multiple nodes.

Summary: The High-Speed Relay