Impressions on Tintri T880

Whoa, It's been months since I last updated my blog. And, our migration project was delayed by many reasons, but it's now going on full speed, so finally I have some experiences on how Tintri performs on real life.

So, we now have about 240 virtual machines running on our T880. And this is how it looks on overall level. Of course, this is just a short sneak, but it looks quite good.

But to be honest, it really should look like that. As you can see in T880 specs, it has 8,8 TB of flash. So, almost all the data still fit's in to that flash.

So how much data do we have in out Tintri now?

Space allocated to VMs: ~37TB
Since all is thin provisioned, logical space consumption is 19TB
And when compression of Tintri is in action, real used space is 9,7TB.

That's actually pretty nice, right?

Space savings

What about latency and flash hit ratio, one points that Tintri uses on their marketing quite heavily. Well, as said, there is only 9.7TB of actual data, so flash hit ratio should be quite good, right? And, it has been, when checking it, it usually always is between 98-100%. But on some occasions, it can be way less, here is a screenshot of last 7 days.

Flash hit ratio

But, what happened? Tintri does autotiering, and keeps hot blocks on flash. And also everything is written first to SSD, then to HDD. And what happened on that timeframe, flash hit ratio was ~50% at it's worst?

What about lantency? Tintri says that it should stay under 1 ms. And, almost all the time it does, but let's have a look of out 7-day graph. Blue is storage latency, so you can see that latency is not caused by hosts or network.

What happened, did Tintri fail?

Well, these two pictures explains a little, first we have IOPS graph (blue is write, yellow is read):

IOPS

And then we have throughput graphs:

So we can see really heavy activity.

And reason for this all is: Full backups. On normal activity, latency and flash hit ratio stays on really good level, but it's quite understandable, that while running full backup, data must be read from HDD, and latency starts to look similar that we see traditional storage systems with HDD disks.

So, have we been happy with this? Yes, we have had no problems since day one.

And our users? Well, we have been migrating servers from really old legacy environment, with EOL hosts and and years old, mid range (actually still quite well performing) storage system.

So when VM's get to new hosts and run on top of Tintri, it's of course a huge improvement on performance, and all the feedback that we have been getting from our users, has been positive.

When we get our migrations done (VM count will probably be something between 500-600), and we have even more load, I'll update these statistics. I'm quite curious to see that will Tintri handle the load, as it should. So far, it has done all that it has promised.