backtop


Print 9 comment(s) - last by .. on Sep 1 at 8:50 AM

Poll data comes in on the idea of virtualization

In a recent poll of nearly 2,000 DailyTech readers, worries about performance seemed to be the biggest potential objection to broader adoption of virtualization technology. Of the 1,998 respondents, 37 percent said they agreed with the statement that “virtualizing slows everything down.” Smaller numbers cited lack of redundancy (20 percent), complexity (12 percent), and cost (9 percent).

More than a fifth of the survey respondents – 439 readers, or 22 percent of the survey group – felt that there were no major drawbacks to virtualization technology, voicing their support for the statement, “Virtualization just rocks. Enough said.”

Several readers were vocal in their disagreement with the perception that virtualization can trigger performance issues. “Having worked with virtualization, I can say that performance should be `almost’ a non-issue with virtualization nowadays,” reader solgae1784 wrote in a comment posted to the DailyTech poll webpage.

“I say `almost’ because there are a few applications that (are) not suitable for virtualization,” solgae1784 added, noting that in such cases “you will see `near native’ performance due to overhead - even if the overhead is supposedly very small.”

DailyTech reader Mjello warned that running SQL on a virtualized server can be problematic. “If you have any sort of SQL on (the server), I’d urge you to not trust VM. If the server actually fails completely, hot migration (won’t) work,” Mjello wrote. “VM failover only works fully as long as the machine has something to migrate from. Otherwise it’s like pulling the plug on a normal server and then booting it again. Not a wise thing to do with a SQL under load.”

Lack of redundancy is a nonissue with virtualization, according to reader PorreKaj. “Virtualization can be expanded over several machines. For example, we have a little blade center with four blades running about 12 servers,” PorreKaj wrote. “If one blade blows up, WMware will just automatically assign the virtual servers to the other blades instantly - without interrupting the user.”

While DailyTech reader solgae1784 was generally bullish on virtualization, he did offer the advice that users should think carefully before choosing which CPU to place inside their virtualized machines. “Be warned that newer CPUs are much better for virtualization than the older CPUs,”  according to solgae1784. “The bottom line is, if you're going for virtualization, you may need to buy new servers. Make sure you do your homework to see if the initial investment will pay off in the long run. Then again, your servers may be due for a hardware refresh anyways.”



Comments     Threshold


This article is over a month old, voting and posting comments is disabled

RE: Ah, virtualization...
By MatthiasF on 8/26/2009 6:24:51 PM , Rating: 1
quote:
It's true that hot migration (VMotion in VMware, Live Migration in Hyper-V) requires a running source host to transfer the VM from.


Not really. Some current techniques involve peer host systems keeping a redundant copy of VMs in RAM, differenced from the SAN at intervals. So, if one host fails, another host would have a somewhat current copy already loaded and ready to run. Not as complicated as VMotion, but it works for many webserver, email and transactional scenarios.

quote:
Virtualization in combination with clustering provides greater availability of resources than clustering alone.


Just to clarify, the only way this would be helpful is if the clustering was being done among virtual machines, which would be a pretty big performance hit. This is completely unnecessary with current technology and in most situations.

quote:
You choose RAID 5 to allow for capacity and data protection


There is no one with any experience building a large server would pick RAID 5. The performance hit starts to make it unwieldy after 5-7 drives. Any cluster or virtualization setup will be using RAID 10, which provides a lot more IOPS and space for the cost. So, let's stick with RAID 10.

Secondly, SAS drives are only meant for space-constrained high-performance scenarios and not SANs. You get more bang for your buck buying enterprise grade SATA than SAS.

Thirdly, you don't build virtualized situations to run right off SANs, since this would create a bottleneck. Virtualized farms will include servers with independant storage that is made redundant by an NFS like inside of VMotion, so all servers share data but have their own storage systems. Only data that must be changed by numerous servers will be placed on an external storage system like a SAN, and performance must be kept in mind when doing this.

Lastly, no one's going to consolidate 20 servers down onto one server (unless it's super simple webservers like at a webhost). At most you'll plan for 5-6 VMs per server on average, less or more depending on what's happening on the VMs.

So, in your scenario there would be four host servers, most likely filled with 8-12 enterprise grade SATA drives (in RAID 10), which among all the servers would provide 4000-6000 IOPS at minimum loads (queue depth 2-4), and more at larger loads (higher command queue depths). Eight is a typical queue depth nowadays and the latest enterprise grade SATA drives can handle 120 IOPS each at this depth.

As far as your 300 IOPS per server number, I think it seems a bit high. Typical planning IOPS numbers by server would make this hypothetical server quite large. For instance, file servers are typically planned at 0.75 IOPS per user, email servers at 0.30, database at 0.90. So, your example would be a 400 user file server, 1000 user email server or 333 user database server.

Webservers are a bit harder to plan if the site is dynamic, but a static file webserver (or FTP) is typically planned at 0.15 IOPS per user.

Anyway, virtualization is not as performance intensive as you make it if you plan it well.

quote:
Besides, with less hardware you should have less to failure to worry about.


I agree. The redundancy concerns aren't realistic. In most scenarios, there is far more redundancy after virtualizing than before.

I also agree with you on the administration. Not only is it easier to handle a crisis, but also easier to do upgrades. When you want to add more or update hardware, you just install the host setup and shuffle duties around between the servers without any downtime in most cases.

quote:
I'll admit that virtualization does not always lower costs.


Virtualization in all viable scenarios will lower costs. Whether it's the large company going from 700 servers to 80, or just a small company with 4-5 servers that wants to be able to move a servers OS from one machine to the other quickly and easily in case of failures, it will always save time and money when implemented well.

The only time virtualization looks too expensive for your use, is when you're trying to use it inappropriately.


RE: Ah, virtualization...
By amanojaku on 8/26/2009 9:08:16 PM , Rating: 2
You raise a few good points, but there a few things I need to clarify.

quote:
Some current techniques involve peer host systems keeping a redundant copy of VMs in RAM, differenced from the SAN at intervals. So, if one host fails, another host would have a somewhat current copy already loaded and ready to run. Not as complicated as VMotion,
VMware calls this FT, or fault tolerance, and it is much more complicated than VMotion. The CPU instructions are mirrored on each VM. If the primary VM fails the secondary picks up at the point the first left off. It's more responsive than traditional clustering, but also more wasteful as it "runs" on the secondary host.
quote:
the only way this would be helpful is if the clustering was being done among virtual machines, which would be a pretty big performance hit.
Not true. Clusters perform very well in VMs. The problem is the cluster software heartbeats, or their sensitivity. Clusters don't like packet latency, and VMs can introduce latency. It's recommended that the number of nodes in a cluster be small, but many clients report success with cluster sizes of a few dozen nodes. With no performance penalty. The trick is building the proper infrastructure.
quote:
There is no one with any experience building a large server would pick RAID 5
My original post was not meant to be a SAN configuration tutorial. It was meant to illustrate the process for determining the bare minimum drives for a LUN. RAID 5 requires the least number of drives compared to the other levels (technically, RAID 0 does, but we never talk about RAID 0.)
quote:
Lastly, no one's going to consolidate 20 servers down onto one server (unless it's super simple webservers like at a webhost). At most you'll plan for 5-6 VMs per server on average, less or more depending on what's happening on the VMs.
Says who? With quad cores and quad sockets you can get 16 cores in a server. That's enough to run 48 single CPU VMs, 24 dual CPU VMs, and four quad CPU VMs with no performance penalty. The typical dual socket quad core server runs 16-20 VMs on average, with a mixture of single and dual core VMs. One client runs 300 VMs on 16 blades in a single chassis, with clusters spread across four blade chassis.
quote:
Secondly, SAS drives are only meant for space-constrained high-performance scenarios and not SANs. You get more bang for your buck buying enterprise grade SATA than SAS.
WTF are you talking about? SAS drives are standard in all servers, and have replaced SCSI drives in SANs, as well. FC is slowly loosing market share to SAS. And there is no enterprise SATA drive that compares to SAS, not in terms of performance or reliability.
quote:
Thirdly, you don't build virtualized situations to run right off SANs, since this would create a bottleneck.
WHAT?!? Ask any storage administrator and he or she will tell you a SAN is THE highest performing storage solution to date, and easier to manage than internal storage. Any bottleneck is the result of poor planning or tight wallets. A key point is that a SAN can support FC, iSCSI or NFS, and some do all simultaneously. In other words, you will always use a SAN, or suffer the consequences of poor performance and limited data mobility.
quote:
As far as your 300 IOPS per server number, I think it seems a bit high. Typical planning IOPS numbers by server would make this hypothetical server quite large. For instance, file servers are typically planned at 0.75 IOPS per user, email servers at 0.30, database at 0.90. So, your example would be a 400 user file server, 1000 user email server or 333 user database server.
That all depends on the company. Anyway, I was presenting an example. Lord 666 figured out that you need to calculate your IOPS before virtualizing.
quote:
Virtualization in all viable scenarios will lower costs. Whether it's the large company going from 700 servers to 80, or just a small company with 4-5 servers that wants to be able to move a servers OS from one machine to the other quickly and easily in case of failures, it will always save time and money when implemented well.
Maybe, maybe not. A business that starts with 5 physical servers that cost $3K each will need at least two virtual host platforms at $10K that can run at least 5 VMs. Then you need shared storage, so you go with a low end SAN. The client doesn't have a SAN, having used local storage all this time, so you pick out something that runs $15K (they go as low as $3K, but you're sacrificing storage and performance.) We haven't even bought virtualization software yet, and we're already spending $10K more to build the infrastructure. There are definitely other benefits to virtualization, but each case has to be evaluated. I know, I do this for a living.


RE: Ah, virtualization...
By MatthiasF on 8/26/2009 10:13:25 PM , Rating: 2
quote:
VMware calls this FT, or fault tolerance, and it is much more complicated than VMotion. The CPU instructions are mirrored on each VM. If the primary VM fails the secondary picks up at the point the first left off. It's more responsive than traditional clustering, but also more wasteful as it "runs" on the secondary host.


No, I'm speaking more to Xen implimentations with centralized management of the hosts. The linux host will be assigned several offline VM's for redundancy that it will keep loaded in some capacity inside of RAM and kept updated over the NFS, so if it had to go online it would be able to pickup the tasks the VM had within a minute or two. While not being as quick a fallover as fault tolerance, it's inexpensive and less a performance hindrance (with the availability of cheap RAM).

quote:
Says who? With quad cores and quad sockets you can get 16 cores in a server. That's enough to run 48 single CPU VMs, 24 dual CPU VMs, and four quad CPU VMs with no performance penalty. The typical dual socket quad core server runs 16-20 VMs on average, with a mixture of single and dual core VMs. One client runs 300 VMs on 16 blades in a single chassis, with clusters spread across four blade chassis.


I highly doubt that in most virtualization environments buying such overpowered machines would be cost effective, nor take advantage of the redundancy in virtualization. Not every situation that virtualization would be helpful involves such performance or redundancy requirements that I'm sure you had to meet in that big client's situation.

quote:
WTF are you talking about? SAS drives are standard in all servers, and have replaced SCSI drives in SANs, as well. FC is slowly loosing market share to SAS. And there is no enterprise SATA drive that compares to SAS, not in terms of performance or reliability.


Performance, no, because they're still using the "SCSI is for servers" mantra of old to sell you overpriced hard drives and not selling high-speed enterprise SATA. Reliability, yes, MTBF are nearly the same for both.

For instance, Seagate's Savvio SAS line has MTBF of 1.6 million hours to Western Digitals RE3 Enterprise SATA line of 1.2 million hours.

So, why spend three times as much for a hard drive that's only a third faster, a third higher MTBF but a third smaller? Exactly as I said, for space-constrained high-performance scenarios. Maybe most of your installations needed it but it's not necessary for the majority of the whole.

quote:
WHAT?!? Ask any storage administrator and he or she will tell you a SAN is THE highest performing storage solution to date, and easier to manage than internal storage. Any bottleneck is the result of poor planning or tight wallets. A key point is that a SAN can support FC, iSCSI or NFS, and some do all simultaneously. In other words, you will always use a SAN, or suffer the consequences of poor performance and limited data mobility.


While today's SAN equipment might be fast, they just aren't necessary anymore for the majority of today's virtualization designs. In fact, several major second generation virtualization initiatives (at big webhosts, CDNs, and Fortune 500 companies) have gone to more independent systems that coordinate without the need for tight centralization. This increases redundancy of the systems, while also bringing costs down greatly at install as well as expansions while also increasing performance.

No longer do they need to spend big bucks on Fiber channel networks on parallel to Ethernet, instead moving to 10Gbe and bringing both networks together.

Most large cloud networks use this method now, including Google.

quote:
That all depends on the company. Anyway, I was presenting an example. Lord 666 figured out that you need to calculate your IOPS before virtualizing.


Always best to do your homework. Probably a good idea to record for at least a month, since some business practices aren't daily, and take an average. Some IT departments record long term data like IOPS.

quote:
Maybe, maybe not. A business that starts with 5 physical servers that cost $3K each will need at least two virtual host platforms at $10K that can run at least 5 VMs. Then you need shared storage, so you go with a low end SAN. The client doesn't have a SAN, having used local storage all this time, so you pick out something that runs $15K (they go as low as $3K, but you're sacrificing storage and performance.) We haven't even bought virtualization software yet, and we're already spending $10K more to build the infrastructure.


I doubt someone running only 5 VMs would bother with a SAN or even buying commercial virtualization software. This lil niche is left to cheap solutions (like Microsoft's Hyper-V) or open source (like Xen) and regular servers if necessary.

Only $30 for Microsoft's Hyper-V server, with the guest license built into the Server 2008 licensing (excluding Web server and Itanium editions).

Xen is free, albeit you'll mostly likely need some knowledge of Linux if you don't buy a pre-packaged solution. I've seen some ISO's setup with Xen in several configurations for free online. Still, probably best to have Linux experience (just like Windows experience is a plus for Microsoft's solution).

quote:
There are definitely other benefits to virtualization, but each case has to be evaluated. I know, I do this for a living.


Yah, there are a lot of us, albeit I've found a lot of early adopter installations that were very poorly designed. The field is still young, so it's good to argue over semantics to help refine things.


"Game reviewers fought each other to write the most glowing coverage possible for the powerhouse Sony, MS systems. Reviewers flipped coins to see who would review the Nintendo Wii. The losers got stuck with the job." -- Andy Marken

Related Articles













botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki