Can PCIe Flash Cut Oracle Redo Log Sync Waits? 4K Log Test Results

Oracle Recommendation on Redo Log Placement

Oracle’s MOS document 857576.1 advises placing online redo logs on SSDs to minimise redo‑log latency and improve the performance of synchronous writes ( log file sync). The same guidance warns that SSD write‑peak behaviour can increase log file sync waits, recommends avoiding RAID‑5/6, and, except for engineered systems (Exadata, SuperCluster, Oracle Database Appliance), suggests not using SSDs for redo logs.

Sector Size and 4K Online Redo Logs

Modern storage devices expose 4 KB physical sectors. Two modes exist:

Native 4 KB mode – logical and physical block sizes are both 4096 bytes. The operating system and applications must support 4 KB I/O.

Emulation (512e) mode – physical blocks remain 4 KB but the logical block size presented to the OS is 512 bytes. Reads of 512 bytes cause a full 4 KB physical read, and writes require a read‑modify‑write cycle, leading to partial I/O and higher latency.

Oracle datafiles use 8 KB blocks and control files use 16 KB blocks, so they are not affected by partial I/O. Online redo logs, however, default to 512‑byte blocks and can suffer from partial I/O on 512e storage.

Since Oracle 11gR2 the redo‑log block size can be set to 512, 1024, or 4096 bytes. Example to create a 4 KB redo log group:

ALTER DATABASE ADD LOGFILE GROUP 5 SIZE 100M BLOCKSIZE 4096;

If the underlying storage reports a 512‑byte sector, Oracle will raise ORA‑01378. The hidden initialization parameter _DISK_SECTOR_SIZE_OVERRIDE can be set to TRUE to bypass the check:

ALTER SYSTEM SET "_DISK_SECTOR_SIZE_OVERRIDE"="TRUE";

Test Environment

Hardware : Enterprise‑grade PCIe flash cards (MLC) and SAS HDDs.

Software : Oracle Database 12.1.0.2, HammerDB TPCC benchmark, 5 000 warehouses of test data, PDB ANIQNG.

Test‑bed photographs:

Testing Procedure

Create an AWR snapshot before the workload:

EXECUTE DBMS_WORKLOAD_REPOSITORY.CREATE_SNAPSHOT();

Run the HammerDB TPCC workload with a defined number of virtual users (first 20, then 120) against each storage configuration.

After the run, generate an AWR report: @?/rdbms/admin/awrrpt.sql Extract the “Load Profile” and “Top 10 Foreground Events by Total Wait Time” sections for analysis.

Results: Flash vs SAS

When the redo logs were placed on PCIe flash (4 KB blocksize) the log file sync wait time fell from 63.6 % of DB time to 18.5 %, and the transactions‑per‑minute (TPM) metric increased noticeably. With SAS disks the TPM curve was less stable and never reached the same peak even at 120 virtual users.

Conclusion

Empirical testing confirms that, for Oracle databases, using enterprise‑grade PCIe flash cards with 4 KB online redo logs provides a clear performance advantage. The 4 KB blocksize eliminates partial I/O on 512e storage, and the flash media’s write‑peak characteristics do not degrade log file sync latency, making flash a viable choice for high‑throughput redo logging.