Best Practices - which domain types should be used to run applications

This post is one of a series of "best practices" notes for Oracle VM Server for SPARC (formerly named Logical Domains)

One question that frequently comes up is "which types of domain should I use to run applications?" There used to be a simple answer in most cases: "only run applications in guest domains", but enhancements to T-series servers, Oracle VM Server for SPARC and the advent of SPARC SuperCluster have made this question more interesting and worth qualifying differently. This article reviews the relevant concepts and provides suggestions on where to deploy applications in a logical domains environment.

Review: division of labor and types of domain

Oracle VM Server for SPARC offloads many functions from the hypervisor to domains (also called virtual machines). This is a modern alternative to using a "thick" hypervisor that provides all virtualization functions, as in traditional VM designs, This permits a simpler hypervisor design, which enhances reliability, and security. It also reduces single points of failure by assigning responsibilities to multiple system components, which further improves reliability and security. In this architecture, management and I/O functionality are provided within domains.

Oracle VM Server for SPARC does this by defining the following types of domain, each with their own roles:

  • Control domain - management control point for the server, used to configure domains and manage resources. It is the first domain to boot on a power-up, is an I/O domain, and is usually a service domain as well.
  • I/O domain - has been assigned physical I/O devices: a PCIe root complex, a PCI device, or a SR-IOV (single-root I/O Virtualization) function. It has native performance and functionality for the devices it owns, unmediated by any virtualization layer.
  • Service domain - provides virtual network and disk devices to guest domains.
  • Guest domain - a domain whose devices are all virtual rather than physical: virtual network and disk devices provided by one or more service domains. In common practice, this is where applications are run.

Typical deployment

A service domain is generally also an I/O domain: otherwise it wouldn't have access to physical device "backends" to offer to its clients. Similarly, an I/O domain is also typically a service domain in order to leverage the available PCI busses. Control domains must be I/O domains, because they boot up first on the server and require physical I/O. It's typical for the control domain to also be a service domain too so it doesn't "waste" the I/O resources it uses.

A simple configuration consists of a control domain, which is also the one I/O and service domain, and some number of guest domains using virtual I/O. In production, customers typically use multiple domains with I/O and service roles to eliminate single points of failure: guest domains have virtual disk and virtual devices provisioned from more than one service domain, so failure of a service domain or I/O path or device doesn't result in an application outage. This is also used for "rolling upgrades" in which service domains are upgraded one at a time while their guests continue to operate without disruption. (It should be noted that resiliency to I/O device failures can also be provided by the single control domain, using multi-path I/O)

In this type of deployment, control, I/O, and service domains are used for virtualization infrastructure, while applications run in guest domains.

Changing application deployment patterns

The above model has been widely and successfully used, but more configuration options are available now. Servers got bigger than the original T2000 class machines with 2 I/O busses, so there is more I/O capacity that can be used for applications. Increased T-series server capacity made it attractive to run more vertical applications, such as databases, with higher resource requirements than the "light" applications originally seen. This made it attractive to run applications in I/O domains so they could get bare-metal native I/O performance.

This is leveraged by the SPARC SuperCluster engineered system, announced a year ago at Oracle OpenWorld. In SPARC SuperCluster, I/O domains are used for high performance applications, with native I/O performance for disk and network and optimized access to the Infiniband fabric.

Another technical enhancement is the introduction of Direct I/O (DIO) and Single Root I/O Virtualization (SR-IOV), which make it possible to give domains direct connections and native I/O performance for selected I/O devices. A domain with either a DIO or SR-IOV device is an I/O domain.

In summary: not all I/O domains own PCI complexes, and there are increasingly more I/O domains that are not service domains. They use their I/O connectivity for performance for their own applications.

However, there are some limitations and considerations: at this time, a domain using physical I/O cannot be live-migrated to another server. There is also a need to plan for security and introducing unneeded dependencies: if an I/O domain is also a service domain providing virtual I/O to guests, it has the ability to affect the correct operation of its client guest domains. This is even more relevant for the control domain. where the ldm has to be protected from unauthorized (or even mistaken) use that would affect other domains. As a general rule, running applications in the service domain or the control domain should be avoided. For reference, an excellent guide to secure deployment of domains by Stefan Hinker is at Secure Deployment of Oracle VM Server for SPARC.

To recap:

  • Guest domains with virtual I/O still provide the greatest operational flexibility, including features like live migration.
  • I/O domains can be used for applications with high performance requirements. This is used to great effect in SPARC SuperCluster and in general T4 deployments. Direct I/O (DIO) and Single Root I/O Virtualization (SR-IOV) make this more attractive by giving direct I/O access to more domains.
  • Service domains should in general not be used for applications, because compromised security in the domain, or an outage, can affect other domains that depend on it. This concern can be mitigated by providing guests' their virtual I/O from more than one service domain, so an interruption of service in the service domain does not cause an application outage.
  • The control domain should in general not be used to run applications, for the same reason. SPARC SuperCluster use the control domain for applications, but it is an exception: it's not a general purpose environment; it's an engineered system with specifically configured applications and optimization for optimal performance.

These are recommended "best practices" based on conversations with a number of Oracle architects. Keep in mind that "one size does not fit all", so you should evaluate these practices in the context of your own requirements.

Summary

Higher capacity T-series servers have made it more attractive to use them for applications with high resource requirements. New deployment models permit native I/O performance for demanding applications by running them in I/O domains with direct access to their devices. This is leveraged in SPARC SuperCluster, and can be leveraged in T-series servers to provision high-performance applications running in domains. Carefully planned, this can be used to provide higher performance for critical applications.

Comments:

Great Post. Are there any Whitepapers that detail the performance impact on using Virtualised I/O vs Native (Network and especially SAN Disk). My interest is related to running a high performance Oracle database within a LDOM.

Posted by chris horne on October 17, 2012 at 05:34 PM MST #

Thanks, Chris - I'm glad you liked!

I don't know of specific whitepapers on this, because it winds up becoming an "it depends on your configuration and workload with wide differences in reported results", which makes it hard to state rules of thumb that say "if you do virtual I/O it costs you x%". I recognize how desirable such figures would be (it would make things a lot simpler for me for sure) but it gets complicated in a hurry with many factors that affect performance.

Guidance I can offer: disk performance is much affected by backend choice, with best performance from giving the LUNs to the guest, rather than mounting an FS in the control domain and creating virtual disks based on files in that. If you do use a ZFS filesystem to host virtual disks, then use zvols rather than flat files for the virtual disks. Don't put all the I/O on a LUN - it's better to have multiple vdisks to avoid queuing or saturating a backend. When optimal performance is needed, consider running the DB in an I/O domain, at which point it will be native performance. I hope that's helpful.

regards, Jeff

Posted by Jeff Savit on October 18, 2012 at 10:54 AM MST #

Post a Comment:
Comments are closed for this entry.
About

jsavit

Search

Categories
Archives
« April 2014
SunMonTueWedThuFriSat
  
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
   
       
Today