During testing, one thing that struck me was the sound of the drives, once they were loaded with information and being tested. The 4k random read is one of the hardest types of file access for any storage solution, especially Hard Disk Drives. At the beginning of each test, sitting in the room I could hear the sound of the drives chattering away as they accessed data when I first started the IOmeter benches. For these being low power drives spinning at 5400 rpm, they were making quite the noise!
As I was curious, I ran a quick test with two Caviar Blacks that I have in another machine, with the same type of file access. They are definitely much quicker drives individually running at 7200 RPM, however, they were also quite a bit louder when they were up and running. You could only imagine the sound of twenty or so 12,000 rpm drives going at it, the sound would be absolutely tremendous!
Why bring up the noise? Well, I began to notice that after a few hours, things became noticeably quieter. At first I chalked it up to me just getting used to the noise, and it becoming somewhat of a ‘elevator music’ that I wasn’t even noticing. Sure enough though, each time I restarted a new run, the noise level would go back up.
The reason the drives were becoming so much quieter is that they were being offloaded of that much of the load once CacheCade had kicked in and was running full force. I am running very limited tests, so I’m sure in data centers getting tens of thousands of hits you wouldn’t even notice the difference in noise levels. Here in my quiet room I sure did though. For there to be that much of a difference just audibly, there is definitely some serious off-loading of work from these drives by the time the benchmark is done and the hot data is cached.
The tester in me decided to do a little ‘informal’ testing, and when I began the next series of tests I setup a little side-test. When I began the ‘Multiple-Zoned” testing, as soon as I loaded the IOmeter test, when the drives were uncached and under full load, I copied a 8 GB folder to the array, multiple times.
Then, after the run had been going for a few hours, and the drives had quieted a bit, I would copy the same 8 gb file folder again, multiple times, to the array.
Using some quick math, I came up with the average copy time of the file folder when the solution wasn’t running cached at 1 min, 40 seconds.
Once the array was well-cached onto the SSDs, the average time of the file transfer dropped to 1 minute and 10 seconds!
That of course, is just an informal testing and some quick number crunching, but a thirty percent increase in write speed was very impressive, and shows that there definitely is more to the performance increase than just the more noticeable read speed performance. Offloading the drives off of the Hot Data when they are under full load definitely will help to increase the performance of your base layer of storage in this type of array.
The only negative aspect of CacheCade that I could gather is that each time the machine is rebooted; the caching has to begin anew. So, if you are using a volume for a few weeks, and the SSD caching is complete and optimized, after a reboot unfortunately you will have to begin the caching process anew. For servers in enterprise environments this would not be a large concern, as the ramp time is small, and the up-time on those servers is going to be as long as possible. For me, personally, I would use this in an operating system environment on my computer if it didn’t have to be re-cached every reboot. The performance gains are just that spectacular. Who wouldn’t want an 8tb drive with the speed of an SSD?
In enterprise environments, price per transaction is one of the key measurements of data center performance. What is the cost per transaction, with all factors included? The amount of power and space required are huge parts of overhead for any data center. And the more space required, coupled with the amount of heat that the components generate, is going to directly affect the cost of cooling. Upgrading with SSDs is a win-win as they also do not require much power, and generate next to no heat.
You must also account for the cost of purchasing and installing the equipment itself.
In the end we must ask what all of this means? The big takeaway here is that CacheCade allows for higher performance plus lower cost and power per transaction.
The price for the upgrade itself is very small considering that most of the infrastructure, in most cases, is already there. There is no need to do massive restructuring, and you can still gain some very large performance gains. One server can essentially begin to do the work of several, just by a topical upgrade.
Purchasing CacheCade is simple, provided that you already own one of the LSI controllers that are compatible with it. For the 9260 and 9265, for example, all you do is purchase a software ‘key’, that you install using a very simple wizard inside of the operating system. It really doesn’t get much easier, then all that is left to do is install the SSDs, and you are a few clicks away from some impressive gains.
More exciting, is that by the end of this year, LSI will be offering this solution with write-caching as well! This will truly be revolutionary, and we cannot wait to see the results.
In summary, I guess it would be a bit of a quick thing to say “works as advertised”, which it does. What is advertised is nothing short of spectacular, and that is exactly what it delivers.
We would like to thank all of the sponsors that made this happen. LSI for providing the 9265-8i for testing, Crucial for providing the C300, Samsung with their wonderful F4EG drives, and Areca for use of their ARC-4036 enclosure, which really makes tasks like this easy.
Page 1- Introduction
Page 2- Nuts and Bolts
Page 3- Maximizing Benefits
Page 4- MLC vs SLC
Page 5- Test Bench and Protocol
Page 6- Single Zone Read
Page 7- Hot-Warm-Cool-Cold
Page 8- Conclusion