字幕列表 影片播放
-
I'd like to invite Cosmos Darwin on stage to give you a demo of what we're doing with Server 2019 and storage class memory.
-
Cosmos! Thanks Jeff, good morning everyone!
-
Organizations around the world are looking to hyper-converged infrastructure for their next wave of datacenter modernization.
-
And there's no better way to go from Hyper-V to hyper-converged infrastructure than with Windows Server 2019,
-
which includes Hyper-V, Storage Spaces Direct, and Software-Defined Networking –
-
everything you need, in one familiar product and license. But more than the affordable price, or the familiar management tools,
-
what customers tell us they really love most is the performance. Two years ago, on this stage, we demoed a single Storage Spaces Direct
-
cluster providing Hyper-V virtual machines with nearly 6.7 million IOPS, or storage operations per second.
-
But that was with Windows Server 2016, and hardware that's now two years old.
-
So this year, at Ignite, it's time to see if we can beat that record.
-
The demo you're about to see is the result of a deep engineering collaboration between Microsoft and our friends at Intel,
-
and it features the very latest in storage innovation: Intel Optane DC persistent memory.
-
Now you can be forgiven you're not familiar with Optane, it's pretty new stuff. An Optane module like this one goes into a
-
DDR4 memory slot but the storage it provides, as Jeff said, is persistent, meaning you can power down, power back up again,
-
and everything that you wrote is still there. And unlike regular RAM,
-
Optane comes in sizes up to 512 GB! And it's fully, natively supported in Windows Server 2019, including with Storage Spaces Direct.
-
The performance you can get out of one of these modules is pretty impressive, but what's more fun is an entire rack of them.
-
It's actually a reference configuration that Microsoft and Intel have been working on together: 12 server nodes,
-
each running Windows Server 2019, and packed with future Intel Xeon Scalable processors, persistent memory, and NVMe.
-
Let's take a look.
-
We'll start in Windows Admin Center, which includes a purpose-built dashboard especially for hyper-converged infrastructure.
-
Here we see an overview of our cluster. We can see we have our 12 server nodes, our 72 drives, and over 300 Hyper-V virtual machines
-
– now, they're turned off right now, but we'll get back to that in just a minute. To round out the tour, we can see we have 14 volumes,
-
which are providing us with nearly 100 TB of usable storage. And as you may have seen from the front photo,
-
this environment's not even fully packed: we have additional PCI slots and drive bays available for expansion.
-
Here are those 12 servers, from Intel. And we can also take a look at the drives, and Windows Admin Center lets us group them by type
-
so we can see that we have 24 persistent memory devices, that are going to be used for cache.
-
If you look closely, you'll notice that these show as 768 GB, how is that possible? Well, if we click in, what we see is that Windows is actually
-
striping these at the memory controller level — this is some, like I said, pretty new stuff. We can also see,
-
back on the drives list, our 48 NVMe drives that are being used for capacity. And if you think of flash as small, think again,
-
because these NVMe SSDs are 8 TB each! Back on the dashboard, I'd like to introduce you to two charts:
-
the first shows us IOPS, or storage requests, across the entire cluster,
-
and the second shows us IO latency (how long those requests are taking, on average) as measured at the file system level in Windows.
-
This is not device latency.
-
OK, are you guys ready to see what this system can do?
-
Before we get there, uh one more thing, I'm going to need an extra digit... I don't want to give anything away, but... alright.
-
Now we saw that we have these 300 Hyper-V virtual machines that are turned off. When we turn them on,
-
each one is going to start running an open source Microsoft tool called DISKSPD to generate storage load.
-
So, shall we start turning them on? Let's start with just the virtual machines on node 1. We'll bring those online, and what you'll see happen,
-
back in Windows Admin Center – and we can timelapse forward here – is that you'll see the IOPS shoot through the tens and
-
hundreds of thousands to one million! One million 4kB random IOPS serving just the virtual machines on node 1.
-
The system isn't even breaking a sweat. Let's turn on some more virtual machines.
-
Let's turn on the VMs on node 2. Now you'll see a pattern start to form here: these virtual machines are going to request
-
another million IOPS, and they're going to get it: two million IOPS from this cluster!
-
All right, I don't want to drag this out: shall we turn on all the rest of the virtual machines? Let's turn them on, from nodes 3 through 12,
-
in kind of a sweep. Now, what we'll see here, is as these virtual machines request more and more and more storage IO,
-
this hyper-converged cluster is able to satisfy it. Through 3, 4, 5 million IOPS,
-
to our previous record of 6 million IOPS. And continuing onward and upward through 7 million, 8 million, 9 million,
-
here's a big milestone, 10 million IOPS. 11, 12, 13 million IOPS from a single hyper-converged cluster!
-
Folks, this is our new record, and we believe, the industry record for any hyper converged platform. The pace of innovation here–
-
The pace of innovation here is really exciting. In just two years, we can do twice as many IOPS,
-
and last time it took us 16 servers, this time we only needed 12.
-
Thank you Cosmos!
