Flash storage, deployed in solid state drives (SSD) can be 400 times faster than disk. However, SSD is expensive. A cottage industry of companies like Tintri are trying different approaches to using SSD in new ways.
Traditional Storage Arrays implement SSD as read/write cache. Traditional Storage Arrays attempts to predict what will be needed based upon current activity and will move items onto SSD to improve performance. Tintri VMStore takes a different approach. They use SSD as primary storage for “hot” data, and then offloads “cold” data to sata. It determines what data is hot based upon performance information gathered at the VM’s disks. Tintri believes that gathering information at this level provides better efficiency than gathering the information at a LUN level because caching often guesses wrong leaving idle data sitting in the cache. Because Tintri looks at VM’s disks, it is easier to identify I/O bottleknecks.
Since Tintri is determining which data is on SSD versus SATA, they argue that administrators no longer need to be concerned about storage tiers. I argue that unless you have enough Tintri storage to meet your needs, you’ll still contend with storage tiers.
Since SSD is expensive it must be used as efficiently as possible. Tintri accomplishes this by first de-duping data, and then compressing it. Wouldn’t this take a performance hit? Remember, SSD can be 400 times faster when compared to traditional disk – the performance hit would be insignificant compared to the faster drive.
SSD use Multi-Level-Cells to hold data. These cells can be overwritten only 5000 to 10000 times before the wear out. If your cache wears out – who cares? Just replace it. But if your primary storage wears out? You have a big problem! Tintri combats this by trying to efficiently use SSD via dedup and compression; and to add error checking and RAID 6 for when the SSD wears out.
Since SSD can be 400 times faster than disk, you can do some creative inefficient things and still be faster. Such as de-dup, compression, and data integrity checking. With error checking and parity, the system will ‘heal itself’ from disk problems on the fly. It also uses a Non-Volatile NVRAM as a write buffer so that if the Tintri mid-stream, the last write can still be constructed: The hypervisor decides to write data; the data goes to the primary NVRAM; the data is copied to secondary NVRAM; then the hypervisor is told the data was written; Tintri confirms the NVRAM data is complete; finally the NVRAM is written to SDD. Again… you would think this might slow things down, but not when you are working 400 times faster than traditional disk.
Another nice item: Trintri does automatic disk alignment for virtual machines.
Much of the Trintri literature compares how much faster SSD is versus disk, how more efficient their vmware-aware performance monitoring is compared to caching, and how Trintri takes advantage of the faster SSD to add efficiency and redundancy for high availability.
