For those of you who missed the June 18th TechTalk on the design for a 20,000 user environment, missed out. Well, not really. Luckily, we recorded the presentation so you can watch it whenever you desire. As you know, the webinar was based on a reference design for a 70,000 user school district. Links to the materials are as follows:
In addition to the materials, we also had some really great questions during the webinar, which I’ve answered below:
- No Way
- Of Course
- You are Crazy!!!
Just kidding, the questions you are probably more interested in are as follows:
|Q: Can you go over a little more on the RAM requirements that you need for the virtual desktops?
A: Sure. Scalability tests conducted were able to run Windows 7 with 768MB of RAM. However, in an actual real-world implementation, you will most likely need more than that. First, you need to break your users down into different categories (Light, Normal and Power). As you move through these, the users will require more RAM. Based on the student population, high school users will use more apps than middle school or elementary. Thus we need to allocate more RAM for those groups of users.
|Q: Great Webinar – Did you experience latency problems using Streaming Applications
A: If you mean in the launching of applications, yes. Streamed applications will take longer to start than installed applications. What we did to help shorten the time (still not the same) was to move the RadeCache for the streamed applications to the virtual desktop’s D Drive, which is persistent across reboots. That means the application cache stays put and is reused for subsequent launches. Of course it still isn’t as fast as installed because streamed applications also validate the application cache is correct.
|Q: how did you design HA for PXE services
A: You will need redundancy in your DHCP environment. Unfortunately, DHCP can only give you 1 address for the TFTP server. If we integrate NetScaler VPX into the environment, we can perform load balancing for the TFTP server. So with 1 IP address we automatically balance across X number of TFTP servers. And NetScaler is smart enough to know when TFTP is down and it will not forward requests onto i
|Q: Virtualization of PVS. How many workloads is the cutover between PVS virtual and PVS physical. I guess 2000 desktops = physical PVS , @40 XenApp servers = Virtual PVS is ok? Can you comment ..ta
A: No good answer, unfortunately. Can you virtualize PVS… Yes. Do we recommend it? Not for large, enterprise deployments. Most people agree that server virtualization makes sense for workloads that do not fully consume a system resources. PVS does not fit as it will utilize your NIC to the fullest. Putting PVS on the hypervisor offers little benefit. However, for small deployments, 200 virtual desktops, you won’t fully utilize the NIC and might be able to consolidate some servers
|Q: How can you figure out network overhead on your environment using PVS streaming to XenDesktop host
A: Testing 🙂 PVS is bursty. You only get traffic when you need more of the vDisk. When you do your pilot, you can look to see how much network traffic PVS is generating for a XenDesktop virtual desktop. Some guidelines are as follows: 1 Gbps NIC should be able to support 500 virtual desktop streams. Booting Windows 7 requires roughly 200-230 MB of data across the wire. As you use Office, 300 more MB is transferred (assuming you are using PowerPoint, Word and Excel). Once that data has been transferred, utilization drops until you need more of the vDisk
|Q: what was the plan for fault tolerance on the virtual desktops if using local storage?
A: Users log onto another virtual desktop. If we assume the desktops are throw away machines, then there should be nothing relevant for the user. Their files, data, and personalization should be on a network share and not on the virtual desktop. If the server fails, a user just connects to a new virtual desktop
|Q: What kind of impact do techs like WAAS and Riverbed have on this type of network traffic?
A: Not certain. The challenge technologies will have is HDX is encrypted. In order for these devices to do caching/compression, they must be able to decrypt the traffic. Once the optimization solution can read the HDX packets, it can then compress not only within a single session, but across perform cross-session compression for users within the same office or school. But then before the packets are placed on the WAN, the traffic must be re-encrypted to protect the data.
|Q: You actually assumed no dial-up?!? Nice neighborhood.
A: There will be dial-up, but just not from the schools. Students can dial into their ISPs or connect via DSL or cable modem and get access to the environment remotely. This means we will want to optimize the HDX protocol for low-bandwidth situations, like those dial-up users
|Q: What RAID type for the Virtual Desktop servers?
A: RAID 10 (1+0). We have 8 spindles on each hypervisor that will support the virtual desktops. This allows us to not require the use of a SAN because 8 spindles should allow for enough IOPS to support the expected virtual desktop load. RAID 10 gives us fault tolerance, but without the huge write penalties we would get with RAID 5.