📜 ⬆️ ⬇️

Testing ssd P4800X

Conducted a modest test glands ssd P4800X to write. How many iops to write to what tasks similar to real ones, it gives out. Test results under the cut.

A server with a P4800X piece of hardware was purchased (you can read about it here ). The main task of the server is not loaded yet, and it is very interesting to look at what the ssd disk for $ 2000 at 375Gb is capable of. Measurement was done only iops record.

To begin, I decided to make tests for updating the database. Made the table with 2 columns integer and 2 columns string. Made 40 flows (20 physically kernels) which would pick a DB. In the database itself 40 million lines, each stream of 1 million. It turned out 16170 iops, as seen in the graphs.


')




3 screenshots: iops, cpu-jumps, read-write. For one period of time. The first block is initialization, as 40 million records in the database were created in one stream. The second block is how they were updated in 40 threads.

Then I redid the table into 4 columns integer - I already received 19130 iops. But still far from the claimed 500,000 iops to write.







These plots show a slice of load from the previous test and also a block with initialization and use. Surprisingly, replacing string-> integer gave + 15% performance.

It is curious that in the first and second tests we run into the speed of 400mb / s recording speed.

I was upset by this result, then I tried to record files directly. I tried to make 40 files of 1 Mb each and in 40 threads write something in a random position. Here this piece of iron issued 243110 iops.







I decided not to torment for more than 15 minutes - such a load is very much consumed by the ssd resource.

Curiously, in this case, the total recording speed: 1gb / sec.

Utilization of 100% - that is, it is impossible to squeeze more out of this piece of iron, perhaps we run into the processor. Where did half of the claimed iops go on the record? Maybe I'm testing wrong? Most likely yes. But sorry, this piece of hardware should not drive benchmarks, but solve the business problem. And then the benchmark is similar to the business problem it solves with such characteristics as you see on the charts.

Next, I tried to make measurements for reading - but 125Gb operatives free caches all that is possible. To leave the script with memory or turn off the cache is considered unacceptable, so I calmed down.

I would like to see such tests before buying this piece of iron.

Now they are there - I hope this will help someone.

The benchmark itself is here .

Ready to apply your load scripts on this piece of hardware.

Source: https://habr.com/ru/post/342714/


All Articles