Be Careful of Storage Performance Track

  • Started
  • Last post
  • 0 Responses
  • hinawhy2018

    Excellent performance is an everlasting pursuit for storage system vendors. Vendors use different performance parameters to promote product performance, such as 100% cache hit ratio (maximum IOPS), 256 KB I/O bandwidth, and 100% write latency. All vendors want you to believe that they can provide the highest performance and can easily handle business pressure during peak hours. You may want to know how to identify all-flash arrays that truly have high performance from amongst a huge number of potentially misleading promotions.

    Track 1: 100% cache hit ratio

    When promoting storage performance, many vendors often boast they can achieve “100% cache hit”. This means that all data is stored in memory, instead of being stored permanently. A criterion to determine whether IT architecture is high performance is to check whether it stores the most frequently accessed data in the location that has the fastest response times.

    This essentially means that it is almost impossible to have a 100% catch hit ratio in customer environments. This is because all-flash storage applies to not only different industries but also to varying business models within industries. Therefore, such a claim must be taken with a grain of salt, and customers should thoroughly check their needs against what the product can offer.

    Trick 2: 100% read performance

    Mainstream SSDs offer high performance in data reads, but their write performance is lacking because each time data is written, SSDs erase data from a NAND before writing new data. This process is called “Program/Erase”, which takes 1 ms to 2 ms in MLC/TLC NAND flash, and leads to huge performance differences in data reads and writes.

    image
    Program/Erase in NAND flash

    SSDs often reserve a large amount of over-provisioning space and the “Erase/Program” process is not required every time data is written onto SSDs. Generally speaking, SSD performance suffers more in mixed read and write scenarios than in read-only scenarios. Therefore, real-life performance values must reflect customer business models (such as 8 KB I/O data blocks, read/write ratio 7:3, and RAID 5/6 groups), instead of boasting the value of 100% read performance.

    image
    Data models in typical scenarios

    Trick 3: Performance stability

    Many vendors only promote reference values that are achieved using a specific model and under certain conditions, which does not show their products’ stability.

    The stability of storage performance is critical, especially during service pressure changes and in complex service environments. When choosing storage SLA, customers regard stability as one of the most important indicators, encouraging vendors to produce limited results, which mislead their customers. Stability indicates that response times of 99% of services must remain stable under certain conditions. If services fluctuate, user experience suffers. However, most customers are, unfortunately, unaware of this issue.