Solution provider’s takeaway: Determining the best use of domain roles, relationships and resources in your customer’s Oracle Solaris 10 system requires you to have a depth of knowledge of Logical Domains. Learn what you need to know about Oracle VM Server for SPARC in this chapter excerpt.
Logical Domains (now Oracle VM Server for SPARC) is a virtualization technology that creates SPARC virtual machines, also called domains. This new style of hypervisor permits operation of virtual machines with less overhead than traditional designs by changing the way guests access physical CPU, memory, and I/O resources. It is ideal for consolidating multiple complete Oracle Solaris systems onto a modern powerful, low-cost, energy-efficient SPARC server, especially when the virtualized systems require the capability to have different kernel levels.
The Logical Domains technology is available on systems based on SPARC chip multithreading technology (CMT) processors. These include the Sun SPARC Enterprise T5x20/T5x40 servers, Sun Blade T6320/T6340 server modules, and Sun Fire T1000/T2000 systems. The chip technology is integral to Logical Domains because it leverages the large number of CPU threads available on these servers. At this writing, that number can be as many as 128 threads in a single-rack unit server and as many as 256 threads in a four-rack unit server. Logical Domains is available on all CMT processors without additional license or hardware cost.
3.1 Overview of Logical Domains Features
Logical Domains creates virtual machines, usually called domains. Each appears to have its own SPARC server. A domain has the following resources:
- Network devices
- OpenBoot environment
- Cryptographic accelerators (optional)
The next several sections describe properties of Logical Domains and explain how they are implemented.
Each domain runs its own instance of Oracle Solaris 10 or OpenSolaris with its own accounts, passwords, and patch levels, just as if each had its own separate physical server. Different Solaris patch and update levels run at the same time on the same server without conflict. Some Linux distributions can also run in domains. Logical Domains support was added to the Linux source tree at the 2.6.23 level.
Domains are isolated from one another. Thus each domain is individually and independently started and stopped. As a consequence, a failure in one domain— even a kernel panic or CPU thread failure—has no effect on other domains, just as would be the case for Solaris running on multiple servers.
Oracle Solaris and applications in a domain are highly compatible with Solaris running on a physical server. Solaris has long had a binary compatibility guarantee; this guarantee has been extended to Logical Domains, making no distinction between running as a guest or on bare metal. Solaris functions essentially the same in a domain as on a non-virtualized system.
3.1.3 Real and Virtual CPUs
One of the distinguishing features of Logical Domains compared to other hypervisors is the assignment of CPUs to individual domains. This approach has a dramatic benefit in terms of increasing simplicity and reducing the overhead commonly encountered with hypervisor systems.
Traditional hypervisors time-slice physical CPUs among multiple virtual machines in an effort to provide CPU resources. Time-slicing was necessary because the number of physical CPUs was relatively small compared to the desired number of virtual machines. The hypervisor also intercepted and emulated privileged instructions that would change the shared physical machine’s state (such as interrupt masks, memory maps, and other parts of the system environment), thereby violating the integrity of separation between guests. This process is complex and creates CPU overhead. Context switches between virtual machines can require hundreds or even thousands of clock cycles. Each context switch to a different virtual machine requires purging cache and translation lookaside buffer (TLB) contents because identical virtual memory addresses refer to different physical locations. This scheme increases memory latency until the caches become filled with fresh content, only to be discarded when the next time slice occurs.
In contrast, Logical Domains is designed for and leverages the chip multithreading (CMT) UltraSPARC T1, T2, and T2 Plus processors. These processors provide many CPU threads, also called strands, on a single processor chip. Specifically, the UltraSPARC T1 processor provides 8 processor cores with 4 threads per core, for a total of 32 threads on a single processor. The UltraSPARC T2 and T2 Plus processors provide 8 cores with 8 threads per core, for a total of 64 threads per chip. From the Oracle Solaris perspective, each thread is a CPU. This arrangement creates systems that are rich in dispatchable CPUs, which can be allocated to domains for their exclusive use.
Logical Domains technology assigns each domain its own CPUs, which are used with native performance. This design eliminates the frequent context switches that traditional hypervisors must implement to run multiple guests on a CPU and to intercept privileged operations. Because each domain has dedicated hardware circuitry, a domain can change its state—for example, by enabling or disabling interrupts—without causing a trap and emulation. The assignment of strands to domains can save thousands of context switches per second, especially for workloads with high network or disk I/O activity. Context switching still occurs within a domain when Solaris dispatches different processes onto a CPU, but this is identical to the way Solaris runs on a non-virtualized server.
One mechanism that CMT systems use to enhance processing throughput is detection of a cache miss, followed by a hardware context switch. Modern CPUs use onboard memory called a cache—a very high-speed memory that can be accessed in just a few clock cycles. If the needed data is present in memory but is not in this CPU’s cache, a cache miss occurs and the CPU must wait dozens or hundreds of clock cycles on any system architecture. In essence, the CPU affected by the cache miss stalls until the data is fetched from RAM to cache. On most systems, the CPU sits idle, not performing any useful work. On those systems, switching to a different process would require a software context switch that consumes hundreds or thousands of cycles.
In contrast, CMT processors avoid this idle waiting by switching execution to another CPU strand on the same core. This hardware context switch happens in a single clock cycle because each hardware strand has its own private hardware context. In this way, CMT processors use what is wasted (stall) time on other processors to continue doing useful work.
This feature is highly effective whether Logical Domains are in use or not. Nonetheless, a recommendation for Logical Domains is to reduce cache misses by allocating domains so they do not share per-core L1 caches. The simplest way to do so is to allocate domains with a multiple of the CPU threads per core—for example, in units of 8 threads on T2-based systems. This approach ensures that all domains have CPUs allocated on a core boundary and not shared with another domain. Actual savings depend on the system’s workload, and may be of minor consideration when consolidating old, slow servers with low utilization.
3.2 Logical Domains Implementation
Logical Domains are implemented using a very small hypervisor that resides in firmware and keeps track of the assignment of logical CPUs, RAM locations, and I/O devices to each domain. It also provides logical channels for communication between domains and between domains and the hypervisor.
The Logical Domains hypervisor is intentionally kept as small as possible for simplicity and robustness. Many tasks traditionally performed within a hypervisor kernel (such as the management interface and performing I/O for guests) are offloaded to special-purpose domains, as described in the next section.
This scheme has several benefits. Notably, a small hypervisor is easier to develop, manage, and deliver as part of a firmware solution embedded in the platform, and its tight focus helps security and reliability. This design also adds redundancy: Shifting functions from a monolithic hypervisor to privileged domains insulates the system from a single point of failure. As a result, Logical Domains have a level of resiliency that is not available in traditional hypervisors of the VM/370, z/VM, or VMware ESX style. Also, this design makes it possible to leverage capabilities already available in Oracle Solaris, providing access to features for reliability, performance, scale, diagnostics, development tools, and a large API set. It has proven to be an extremely effective alternative to developing all these features from scratch.
3.2.1 Domain Roles
Domains are used for different roles, and may be used for Logical Domain infrastructure or applications. The control domain is an administrative control point that runs Solaris or OpenSolaris and the Logical Domain Manager services. It has a privileged interface to the hypervisor, and can create, configure, start, stop, and destroy other domains. Service domains provide virtualized disk and network devices for other domains. I/O domains have direct access to physical I/O devices and are typically used as service domains to provide access to these devices. The control domain also is an I/O domain and can be used as a service domain. Applications generally run in guest domains, which are non-I/O domains using virtual devices provided by service domains. The domain structure and the assignment of CPUs are shown in Figure 3.1.
Figure 3.1: Control and Guest Domains
The definition of a domain includes its name, amount of RAM and number of CPUs, its I/O devices, and any optional hardware cryptographic accelerators. Domain definitions are made by using the command-line interface in the control domain, using the Oracle Enterprise Manager Ops Center product, or for the initial configuration, using the Logical Domains Configuration Assistant.
184.108.40.206 Domain Relationships
Each server has exactly one control domain, found on the instance of Solaris that was first installed on the system. It runs Logical Domain Manager services, which are accessed by a command-line interface provided by the ldm command. These Logical Domain Manager services include a “constraint manager” that decides how to assign physical resources to satisfy the specified requirements (the “constraints”) of each domain.
There can be as many I/O domains as there are physical PCI buses on the system. An I/O domain is often used as a service domain to run virtual disk services and virtual network switch services that provide guest domain virtual I/O devices.
Finally, there can be as many guest domains as are needed for applications, subject to the limits associated with the installed capacity of the server. At the time of this writing, the maximum number of domains on a CMT system was 128, including control and service domains, even on servers with 256 threads such as the T5440. While it is possible to run applications in control or service domains, it is highly recommended, for stability reasons, to run applications only in guest domains. Applications that require optimal I/O performance can be run in an I/O domain to avoid virtual I/O overhead, but it is recommended that such an I/O domain not be used as a service domain.
A simple configuration consists of a single control domain also acting as a service domain, and some number of guest domains. A more complex configuration could use redundant service domains to provide failover in case of a domain failure or loss of a path to an I/O device.
3.2.2 Dynamic Reconfiguration
CPUs and virtual I/O devices can be dynamically added to or removed from a Logical Domain without requiring a reboot. An Oracle Solaris instance running in a guest domain can immediately make use of a dynamically added CPU for additional capacity and can also handle the removal of all but one of its CPUs. Virtual disk and network resources can also be nondisruptively added to or removed from a domain, and a guest domain can make use of a newly added virtual disk or network device without a reboot.
3.2.3 Virtual I/O
Logical Domains technology abstracts underlying I/O resources to virtual I/O. It is not always possible to give each domain direct access to a bus, an I/O memory mapping unit (IOMMU), or devices, so Logical Domains provides a virtual I/O (VIO) infrastructure to provide access to these resources.
Virtual network and disk I/O is provided to Logical Domains by service domains. A service domain runs Solaris and usually has direct connections to a PCI bus connected to physical network and disk devices. In that configuration, it is also an I/O domain. Likewise, the control domain is typically configured as a service domain. It is also an I/O domain, because it requires access to I/O buses and devices to boot up.
The virtual I/O framework allows service domains to export virtual network and disk devices to other domains. Guest domains use these devices exactly as if they were dedicated physical resources. Guest domains perform virtual I/O to virtual devices provided by service domains. Service domains then proxy guests’ virtual I/O by performing I/O to back-end devices, which are usually physical devices. Virtual device characteristics are described in detail later in this chapter.
Guest domains have network and device drivers that communicate with I/O domains through Logical Domain Channels (LDCs) provided by the hypervisor. The addition of device drivers that use LDCs rather than physical I/O is one of the areas in which Solaris has been modified to run in a logical domain—, an example of paravirtualization discussed in Chapter 1, “Introduction to Virtualization.” LDCs provide communications channels between guests, and an API for enqueuing and dequeuing messages that contain service requests and responses. Figure 3.2 shows the relationship between guest and service domains and the path of I/O requests and responses.
Figure 3.2: Service Domains Provide Virtual I/O
Shared memory eliminates the overhead associated with copying buffers between domains. The processor’s memory mapping unit (MMU) is used to map shared buffers in physical memory into the address spaces of a guest and an I/O domain. This strategy helps implement virtual I/O efficiently: Instead of copying the results of a disk read from its own memory to a guest domain’s memory, an I/O domain can read directly into a buffer it shares with the guest. This highly secure mechanism is controlled by hypervisor management of memory maps.
I/O domains are designed for high availability. Redundant I/O domains can be set up so that system and guest operation can continue if a path fails, or if an I/O domain fails or is taken down for service. Logical Domains provides virtual disk multipathing, thereby ensuring that a virtual disk can remain accessible even if a service domain fails. Domains can use IP network multipathing (IPMP) for redundant network availability.
3.3 Details of Domain Resources
Logical Domains technology provides flexible assignment of hardware resources to domains, with options for specifying physical resources for a corresponding virtual resource.
3.3.1 Virtual CPUs
As mentioned in the section “Real and Virtual CPUs,” each domain is assigned exclusive use of a number of CPUs, also called threads or strands. Within a domain, these are called virtual CPUs (vCPUs).
The granularity of assignment is a single vCPU. A domain can have from one vCPU up to all the vCPUs on the server. On UltraSPARC T1 systems (T1000 and T2000), the maximum is 8 cores with 4 threads, for a total of 32 vCPUs. On UltraSPARC T2 and T2 Plus systems, the maximum is 8 cores with 8 threads each, for a total of 64 vCPUs per chip. Systems with the T2 Plus chip can have multiple chips per server: The T5140 and T5240 servers have 2 T2 Plus chips for a total of 16 cores and 128 vCPUs, while the T5440 has 4 T2 Plus chips with 32 cores and 256 vCPUs.
Virtual CPUs should be assigned to domains on core boundaries. This strategy prevents “false cache sharing,” which can reduce performance when multiple domains share a CMT core and compete for the same L1 cache. To avoid this problem, vCPU quantities equivalent to entire cores to each domain should be allocated. For example, you should allocate vCPUs in units of 8 vCPUs on T2 and T2 Plus servers. Of course, this tactic may be overkill for some workloads, and administrators need not excessively concern themselves when defining domains to accommodate the light CPU requirements needed to consolidate small, old, or low utilization servers. Figure 3.3 is a simplified diagram of the threads, cores, and caches in a SPARC CMT chip.
The number of CPUs in a domain can be dynamically and nondisruptively changed while the domain is running. Oracle Solaris commands such as vmstat and mpstat can be used within the domain to monitor its CPU utilization, just as on a dedicated server. The ldm list command can be used in the control domain to display each domain’s CPU utilization. A change in the quantity of vCPUs in a running domain takes effect immediately. The number of CPUs can be managed automatically with the Logical Domains Dynamic Resource Manager, which is discussed later in this chapter.
Figure 3.3: CMT Cores, Threads, and Caches
3.3.2 Virtual Network Devices
Guests have one or more virtual network devices connected to virtual Layer 2 network switches provided by service domains. Virtual network devices can be on the same or different virtual switches so as to connect a domain to multiple networks, provide increased availability using IPMP (IP Multipathing), or increase the bandwidth available to a guest domain.
From the guest perspective, virtual network interfaces are named vnetN, where N is an integer starting from 0 for the first virtual network device defined for a domain. In fact, the simplest way to determine if an Oracle Solaris instance is running in a guest domain (specifically, a domain that is not an I/O domain) is to issue the command ifconfig –a and see if the network interfaces are vnet0, vnet1, and so on, rather than real devices like nxge0 or e1000g0. Virtual network devices can be assigned static IP or dynamic IP addresses, just as with physical network devices.
220.127.116.11 MAC Addresses
Every virtual network device has its own MAC address. This is different from Oracle Solaris Containers, where a single MAC address is usually shared by all Containers in a Solaris instance. MAC addresses can be assigned manually or automatically from the reserved address range of 00:14:4F:F8:00:00 to 00:14:4F:FF:FF:FF. The bottom half of the address range is used for automatic assignments; the other 256K addresses can be used for manual assignment.
The Logical Domains manager implements duplicate MAC address detection by sending multicast messages with the address it wants to assign and listening for a possible response from another machine’s Logical Domains manager saying the address is in use. If such a message comes back, it randomly picks another address and tries again. The message time-to-live (TTL) is set to 1, and can be changed by the SMF property ldmd/hops. Recently freed MAC addresses from removed domains are used first, to help prevent DHCP servers from exhausting the number of addresses available.
18.104.22.168 Network Connectivity
Virtual switches are usually assigned to a physical network device, permitting traffic between guest domains and the network segment to which the device is connected. Network traffic between domains on the same virtual switch does not travel to the virtual switch or to the physical network, but rather is implemented by a fast memory-to-memory transfer between source and destination domains using dedicated LDCs. Virtual switches can also be established without a connection to a physical network device, which creates a private secure network not accessible from any other server. Virtual switches can be configured for securely isolated VLANs, and can exploit features such as VLAN tagging and jumbo frames.
22.214.171.124 Hybrid I/O
Network Interface Unit (NIU) Hybrid I/O is an optimization feature available on servers based on the UltraSPARC T2 chip, the T5120 and T5220 servers, and the Sun Blade T6320 server module. It is an exception to the normal Logical Domains virtual I/O model, and provides higher performance network I/O. In hybrid mode, DMA resources for network devices are loaned to a guest domain so it can perform network I/O without going through an I/O domain. In this mode, a network device in a guest domain can transmit unicast traffic to and from external networks at essentially native performance. Multicast traffic, and network traffic to other domains on the same virtual switch are handled as described above.
In current implementations, there are two 10 GbE NIU nxgeN devices per T2- based server. Each can support three hybrid I/O virtual network devices, for a total of six.
3.3.3 Virtual Disk
Service domains can have virtual disk services that export virtual block devices to guest domains. Virtual disks are based on back-end disk resources, which may be physical disks, disk slices, volumes, or files residing in ZFS or UFS file systems. These resources could include any of the following:
- A physical block device (disk or LUN)—for example, /dev/dsk/c1t48d0s2
- A slice of a physical device or LUN—for example, /dev/dsk/c1t48d0s0
- A disk image file residing in UFS or ZFS—for example, /path-to-filename
- A ZFS volume—for example, zfs create -V 100m ldoms/domain/test/zdisk0 creates the back-end /dev/zvol/dsk/ldoms/domain/test/zdisk0
- A volume created by Solaris Volume Manager (SVM) or Veritas Volume Manager (VxVM)
- A CD ROM/DVD or a file containing an ISO image
A virtual disk may be marked as read-only. It also can be made exclusive, meaning that it can be given to only one domain at a time. This setting is available for disks based on physical devices rather than files, but the same effect can be provided for file back-ends by using ZFS clones. The advantages of ZFS—such as advanced mirroring, checksummed data integrity, snapshots, and clones—can be applied to both ZFS volumes and disk image files residing in ZFS. ZFS volumes generally provide better performance, whereas disk image files provide simpler management, including renaming, copying, or transmission to other servers.
In general, the best performance is provided by virtual disks backed by physical disks or LUNs, and the best flexibility is provided by file-based virtual disks or volumes, which can be easily copied, backed up, and, when using ZFS, cloned from a snapshot. Different kinds of disk back-ends can be used in the same domain: The system volume for a domain can use a ZFS or UFS file system back-end, while disks used for databases or other I/O intensive applications can use physical disks.
Redundancy can be provided by using virtual disk multipathing in the guest domain, with the same virtual disk back-end presented to the guest by different service domains. This provides fault tolerance for service domain failure. A timeout interval can be used for I/O failover if the service domain becomes unavailable. The ldm command syntax for creating a virtual volume lets you specify an MPXIO group. The following commands illustrate the process of creating a disk volume back-end served by both a control domain and an alternate service domain:
# ldm add-vdsdev mpgroup=foo \
/path-to-backend-from-primary/domain [email protected]
# ldm add-vdsdev mpgroup=foo \
/path-to-backend-from-alternate/domain [email protected]
# ldm add-vdisk ha-disk [email protected] myguest
Multipathing can also be provided from a single I/O domain with multiplexed I/O (MPXIO), by ensuring that the domain has multiple paths to the same device—for example, two FC-AL HBAs to the same SAN array. You can enable MPxIO in the control domain by running the command stmsboot -e. That command creates a single, but redundant path to the same device. The single device is then configured into the virtual disk service. Perhaps most simply, insulation from a path or media failure can be provided by using a ZFS file pool with mirror or RAID-Z redundancy. These methods offer resiliency in case of a path failure to a device, but do not insulate the system from failure of a service domain.
3.3.4 Console and OpenBoot
Every domain has a console, which is provided by a virtual console concentrator (vcc). The vcc is usually assigned to the control domain, which then runs the Virtual Network Terminal Server daemon (vntsd) service.
By default, the daemon listens for localhost connections using the Telnet protocol, with a different port number being assigned for each domain. A guest domain operator connecting to a domain’s console first logs into the control domain via the ssh command so that no passwords are transmitted in cleartext over the network; the telnet command can then be used to connect to the console.
Optionally, user domain console authorization can be implemented to restrict which users can connect to a domain’s console. Normally, only system and guest domain operators should have login access to a control domain.
3.3.5 Cryptographic Accelerator
The processors in CMT systems are equipped with on-chip hardware cryptographic accelerators that dramatically speed up cryptographic operations. This technique improves security by reducing the CPU consumption needed for encrypted transmissions, and makes it possible to transmit secure traffic at wire speed. Each CMT processor core has its own hardware accelerator unit, making it possible to run multiple concurrent hardware-assisted cryptographic transmissions.
In the T1 processor used on the T1000 and T2000 servers, the accelerator performs modular exponentiation and multiplication, which are normally CPU-intensive portions of cryptographic algorithms. The accelerator, called the Modular Arithmetic Unit (MAU), speeds up public key cryptography (i.e., RSA, DSA, and Diffie-Hellman algorithms).
Although the T2 and T2 Plus chips include this function, the accelerator has additional functionality. This cipher/hash unit accelerates bulk encryption (RC4, DES, 3DES, AES), secure hash (MD5, SHA-1, SHA-256), other public key algorithms (elliptical curve cryptography), and error-checking codes (ECC, CRC32).
At this time, a cryptographic accelerator can be allocated only to domains that have at least one virtual CPU on the same core as the accelerator.
The Logical Domains technology dedicates real memory to each domain, instead of using virtual memory for guest address spaces and swapping them between RAM and disk, as some hypervisors do. This approach limits the number and memory size of domains on a single CMT processor to the amount that fits in RAM, rather than oversubscribing memory and swapping. As a consequence, it eliminates problems such as thrashing and double paging, which are experienced by hypervisors that run virtual machines in virtual memory environments.
RAM can be allocated to a domain in highly granular units—the minimum unit that can be allocated is 4 MB. The memory requirements of a domain running the Oracle Solaris OS are no different from running Solaris on a physical machine. If a workload needs 8 GB of RAM to run efficiently on a dedicated server, it will need the same amount when running in a domain.
3.3.7 Binding Resources to Domains
The Logical Domains administrator uses the ldm command to specify the resources required by each domain: the amount of RAM, the number of CPUs, and so forth. These parameters are sometimes referred to as the domain’s constraints.
A domain that has been defined is said to be inactive until resources are bound to it by the ldm bind command. When this command is issued, the system selects the physical resources required by the domain’s constraints and associates them with the domain. For example, if a domain requires 8 CPUs, the domain manager selects 8 CPUs from the set of online and unassigned CPUs on the system and gives them to the domain.
Until a domain is bound, the sum of the constraints of all domains can exceed the physical resources available on the server. For example, one could define 10 domains, each of which requires 8 CPUs and 8 GB of RAM on a machine with 64 CPUs and 64 GB of RAM. Only the domains whose constraints are met can be bound and started. In this example, the first 8 domains to be bound would boot. Additional domains can be defined for occasional or emergency purposes, such as a disaster recovery domain defined on a server normally used for testing purposes.
Oracle VM Server for SPARC
This excerpt is from the book, ‘Oracle Solaris 10 System Virtualization Essentials’ by Jeff Victor, Jeff Savit, Gary Combs, Simon Hayler, Bob Netherton, published by Pearson/Prentice Hall Professional, Sept. 2010, ISBN 0-13-708188-X, Copyright 2011; for a complete list of contents, please visit the publisher site: www.informit.com/title/013708188X.