Hi
I had one doubt about interpreting the output. Is it worth looking at the ‘trans’ section ? In my case, latencies in ‘trans’ section 80% of requests are greater than 1 ms. But in the ‘device’ section, only 0.90% are more than 1 ms. What’s the significance of trans section in understanding the effectiveness of SSD ?
Short answer : No, its not worth worrying too much about trans histogram if device is good.
Typically, the trans histogram and the device histogram are close. The trans histogram has the overhead of simulating a transaction system over the device I/O. i.e we generate work, put on the queue, dequeue, do the disk I/O, and get the response back. trans histogram includes all these stages. If the queue processing is not fast for whatever reason, the trans histogram will show bad value. If trans histogram is bad and device is good, my guess it that you are hitting system limitations like CPU/memory etc.
Thanks a lot @sunil for your reply. And I am sorry to having missed it.
simulating a transaction system over the device I/O
But actual run’s behavior will be also same, isn’t it ? In actual server benchmark with SSD, my cores are > 50% usage in I/O waiting but SSD has not reached its maximum utilizatiion and IOPS. So, perhaps Aerospike is slow in its RW to SSD Queues as suggested by the ACT also.
You are right. When I said that you need not worry, I am only talking purely from the device point of view. I was assuming that your ACT benchmark machine and your production machine where you will run Aerospike are different. If you are using the production machine and doing the ACT, then yes, the high transaction latency also should be of concern for you. However, the Aerospike software has more tunables for performance than the ACT tool.
The best way is to run the benchmark end-to-end. i.e Try to generate desired aerospike read/write with the planned SSD drives and see if both together are meeting your desired performance or not. If not, we should go back to the drawing board and do the sizing exercise again.