We recently found some performance regressions related to writeback. See http://marc.info/?l=linux-kernel&m=132391033818515&w=2. The test system uses a LSI 1068e card and 12 hard disks attached to it. The workload is a simple FIO test, which runs one thread for each disk to create files. The filesystem is EXT4. Compared to 3.1 kernel, we found both data=order and data=writeback mode have performance regression. data=writeback regression is bigger. The scsi_host->can_queue = 127, where disk queue_depth=64, 64*12 > 127, so the fairness to utilize host can_queue is important for the overall throughput, otherwise some disks will be starved and others run in full speed. In this workload, we found recent I/O less writeback causes the fairness issue and the throughput get dropped. We'd better fix the fairness issue in writeback, but on the other hand, the scsi layer also has fairness issue regarding to the starve list, which exists in old kernel too, but not that severe. Below two patches try to address the scsi issue, the test result is (average of 3 runs, data=writeback mode): 3.1: 1000045 KB/s base: 854016 KB/s base+patch: 958122 KB/s The patches don't fully recover the regression, but give a 12% improvement. Thanks, Shaohua -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
