As many may have seen, I have been running (lots) of benchmarks on different higher-end SSDs.
Looking at IOMeter results the output has many dimensions. For example:
So that is essentially 4 * 6 or 7 (128 max or 256 QD max) sets of results. For each of those I have latency numbers as 21 different buckets (e.g. 0-50us, 1-2s). Even 24 sets of results with 21 buckets = 504 data points which seems like too much.
That does not include throughput, IOPS or other data points either.
Frankly, the latency bit is one of the most exciting aspects of these drives. When I looked into 1 billion records in the five latency buckets > 1s, I only found 21 operations that were in those ~1100 Drive/ Test Profile/ QD/ latency buckets.
Does anyone have thoughts around what data STH should present?
Here is Option 1 - All latency buckets
Option 2 - Consolidate 20ms and >1s
Option 3 - Consolidate further (10-20ms bucket)
Mind you, these figures are aggregated across drives, tests, and queue depths. I am also not finished going through all of the data sets yet but that shows the overall shape and this is not the final graph format. I I just wanted to get a sense of what folks thought on the bucketing. Personally, I like Option 3 but I did want to open this up for discussion.
Looking at IOMeter results the output has many dimensions. For example:
- Database, Web, File Server and Workstation Profiles
- 4-128 or 256 QD using exponential scaling
So that is essentially 4 * 6 or 7 (128 max or 256 QD max) sets of results. For each of those I have latency numbers as 21 different buckets (e.g. 0-50us, 1-2s). Even 24 sets of results with 21 buckets = 504 data points which seems like too much.
That does not include throughput, IOPS or other data points either.
Frankly, the latency bit is one of the most exciting aspects of these drives. When I looked into 1 billion records in the five latency buckets > 1s, I only found 21 operations that were in those ~1100 Drive/ Test Profile/ QD/ latency buckets.
Does anyone have thoughts around what data STH should present?
Here is Option 1 - All latency buckets
Option 2 - Consolidate 20ms and >1s
Option 3 - Consolidate further (10-20ms bucket)
Mind you, these figures are aggregated across drives, tests, and queue depths. I am also not finished going through all of the data sets yet but that shows the overall shape and this is not the final graph format. I I just wanted to get a sense of what folks thought on the bucketing. Personally, I like Option 3 but I did want to open this up for discussion.