So I looked at the throughput of a kv-sync batch. The bytes in throttle are the number of bytes that were in the throttle (no throttling is actually taking place, however the data that would be controlling the throttle is noted) when the earliest TransContext of the block was created (in other words, the # of bytes in the batch aren't added to this). The time is calculated as the difference in when the kv-sync batch was synced and the time that the earliest TransContext in the batch was created. And the bytes is the total number of bytes in the kv-sync batch. The data is generated using bluestore on a hard disk. Here are the plots with work generated by rados bench with varying number of concurrent ops. Please note that the scales on the axes can vary. throughput @ concurrency=16 http://tracker.ceph.com/attachments/download/2707/regbench-m-t16-kv-sync-throughput.pdf throughput @ concurrency=32 http://tracker.ceph.com/attachments/download/2708/regbench-m-t32-kv-sync-throughput.pdf throughput @ concurrency=64 http://tracker.ceph.com/attachments/download/2709/regbench-m-t64-kv-sync-throughput.pdf throughput @ concurrency=128 http://tracker.ceph.com/attachments/download/2710/regbench-m-t128-kv-sync-throughput.pdf throughput @ concurrency=all-overlaid http://tracker.ceph.com/attachments/download/2711/regbench-m-tmulti-kv-sync-throughput.pdf I don't think the data is too surprising. With fewer bytes in the throttle we have some high throughput numbers, but throughput quickly settles into the 400000-500000 bytes/nanosecond range at the peak. The peak then starts to descend as the number of bytes in the throttle increases. Since except for a few stragglers, the most dense area of points is below the 600000 bytes/ns area, here are the same four graphs where we limit the vertical axis to 600000. The scales on the horizontal axes can still vary. throughput @ concurrency=16 http://tracker.ceph.com/attachments/download/2712/regbench-m-t16-kv-sync-throughput-pin.pdf throughput @ concurrency=32 http://tracker.ceph.com/attachments/download/2713/regbench-m-t32-kv-sync-throughput-pin.pdf throughput @ concurrency=64 http://tracker.ceph.com/attachments/download/2714/regbench-m-t64-kv-sync-throughput-pin.pdf throughput @ concurrency=128 http://tracker.ceph.com/attachments/download/2715/regbench-m-t128-kv-sync-throughput-pin.pdf throughput @ concurrency=all-overlaid http://tracker.ceph.com/attachments/download/2716/regbench-m-tmulti-kv-sync-throughput-pin.pdf And the downward trend in throughput as the bytes that would be in the throttle gets larger is even more pronounced. I'm going to fit polynomial curves to this data next. Eric -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
