Rate This Document
Findability
Accuracy
Completeness
Readability

22.0.0.SPC3

The following lists the issue resolved in the Kunpeng BoostKit 22.0.0.SPC3 SDS Global Cache 1.3.15 patch.

Trouble Ticket No.

DTS2022122207239

Symptom

The write performance in the non-full hit scenario with the block size being 128 or 256 KB is much lower than the expected performance (with a fluctuation rate greater than 10%), which does not meet the requirement.

Severity

Major

Cause Analysis

When the block size of wcache is 128 KB or 256 KB, write back is performed by default. The granularity of write data stored in wcache is 4 KB+64 bytes whereas the data flushing granularity is 4 KB. To reduce the number of I/Os, the persistence layer supplements the data flushing granularity to 8 KB, doubling the bandwidth. As a result, the current NVMe drive encounters a bandwidth bottleneck.

Solution

Change the write-through threshold to 128 KB.

Impact

The defect is rectified, and no other impact is imposed.

Test Suggestion

Reserve 3 TB storage space for Ceph and use 16 KB data blocks to perform 7:3 random read/write operations.