I am testing my DC S3500 SSD. I found the write iops and read iops are all OK.
4K bytes read 7K iops,
4K bytes write: 15K iops
512 bytes read 18K iops
512 bytes write 18K iops
Mixed read /write still alright:
random 4k bytes 50% read, 50% write: 6k iops.
But when I tried write & read-back, i.e. random write but immediate read back from same position, I got terrible figure:
random 4k bytes write then read-back: 1000 iops only.
Is there anything wrong for write&read-back?
According to this documentation: http://www.intel.com/content/dam/www/public/us/en/documents/product-specifications/ssd-dc-s3500-spec... Intel® Solid-State Drive DC S3500 Series - Product Specification the advertised performance is tested with 4K 8K only which matches what you have informed if you meant to write 70.000 IOPS for 4K read.
The Intel® SSD DC S3500 Series are not tested with the type of test that you have mention, therefore, we don't have a point of comparison for that.
However, this type of test is expected to be slower. We advise you to do the test based on the specification noted in the document above.
Thank you Aleki.
I did a little more profiling for the read-back after write. Here's what I got:
2000 write N read-back
Finished in- 4,019,333 micro seconds
write: 253,671 (microseconds)
Reading is about 15 times slow than writing if it is a read-back.
My question is, isn't there an internal cache that the reading should be returned immediately?
My industry requires transactional data storage, like database does. Thus a read-back verification is ideal. I guess SSD itself has some kind of verification inside the box. But we still love to see an explicit read-back verification if possible.
Does Intel has some documents or presentation how the data integrity is done or data error is monitored? Thanks.