Flash in a flash – Tintri

We ask Rex Walters, Vice President of Technology at Tintri on his view of the company's Flash value proposition.
Data&StorageAsean: All Flash or Hybrid?
Rex WaltersRex Walters (photo right): Technically, each Tintri VMstore™ storage node physically contains both solid state drives (flash SSDs) as well as hard disk drives (HDDs). However, we use a 100% flash filesystem, with 100% of all metadata residing in flash at all times — we never have to perform an HDD IO to find out where a block resides. Further IO always goes to flash first, and we only move blocks to HDD that we know for certain are “cold" (we count the frequency of accesses and only move the least frequently accessed blocks to HDD). In other words, unlike traditional “hybrid” arrays that use flash as a cache, we assume all blocks need flash performance until we’ve measured otherwise.
Data&StorageAsean: What’s truly Unique about your Flash offering?
Rex Walters: Each Tintri VMstore™ node allocates resources (including NVRAM, cpu, network bandwidth, HDD IOPS etc., not just flash) per virtual volume not per datastore/storage-container/LUN/volume/aggregate/raid-set. Tintri recognizes that modern IT virtualizes applications by default, and that IO workloads differ by virtual volume. A virtualized database has very different IO needs than a virtualized web server, for example. A database presents very different IO to a virtual volume used as a transaction log than it does to one containing indexes or tables. Tintri treats each virtual volume as the fundamental unit of management. Where traditional storage arrays allocate resources and assign policies per physical volume or LUN, Tintri does it per VM and per virtual volume.

This results in far fewer datastores to manage, and far more agile infrastructure for heavily
Data&StorageAsean: Are IOPS everything? What else should people consider when selecting Flash Storage?
Rex Walters: Absolutely not. In fact, IOPS are no longer a very relevant metric for measuring storage performance with solid state storage devices. IOs come in various sizes: an individual 256 KB IO request does more “work” than, say, a 1 KB IO request. This fact was somewhat irrelevant during HDD era when the fastest hard disk drive technology (15K RPM fibre channel drives) could perform roughly 180 random IO requests per second regardless of the block sizes. We learned over the past several decades that the limiting was IOPS and the number of “spindles," not throughput (MB/s).

Solid state storage behaves differently. While an individual SSD can perform orders of magnitude more IOPS than an individual HDD, the number of IOPS roughly halves each time the block size is doubled. The limiting factor with SSDs is really throughput (MB/s) more than IOPS. Specifying IOPS to a solid state array without specifying the block size distribution is meaningless!

Tintri has stolen an idea from Microsoft and uses the concept of “normalized IOPS” (really a proxy for throughput) in our upcoming quality-of-service controls. Normalized IOPS are quantized to 8 KB. A single 1 KB IO is “counted” as one normalized IOP. An 8 KB transaction is also counted as one normalized IOP. But a single 32 KB IO, for example, counts as four normalized IOPS. This allows administrators to specify policies in terms of IOPS as desired, but actually throttle throughput as desired.
Data&StorageAsean: Is there any customer type that your Flash Offering is not suited to?
Rex Walters: Tintri does not support physical workloads (applications running directly on physical servers rather than in a VM). We only support virtual machines and virtual volumes.
Data&StorageAsean: What’s your target market in South Asia
Rex Walters: Tintri currently has an established presence in Japan, Australia, and Singapore and intends to expand throughout all of Asia, including ANZ and ASEAN regions. Tintri is expanding globally, with significant operations in North America and EMEA as well as Asia.

You might also like
Most comment
share us your thought

0 Comment Log in or register to post comments