Contents
Currently Ganeti can be used to easily share a node between multiple virtual instances. While it’s easy to do a completely “best effort” sharing it’s quite harder to completely reserve resources for the use of a particular instance. In particular this has to be done manually for CPUs and disk, is implemented for RAM under Xen, but not under KVM, and there’s no provision for network level QoS.
We want to make it easy to partition a node between machines with exclusive use of hardware resources. While some sharing will anyway need to happen (e.g. for operations that use the host domain, or use resources, like buses, which are unique or very scarce on host systems) we’ll strive to maintain contention at a minimum, but won’t try to avoid all possible sources of it.
exclusive_storage is a new node parameter. When it’s enabled, Ganeti will allocate entire disks to instances. Though it’s possible to think of ways of doing something similar for other storage back-ends, this design targets only plain and drbd. The name is generic enough in case the feature will be extended to other back-ends. The flag value should be homogeneous within a node-group; cluster-verify will report any violation of this condition.
Ganeti will consider each physical volume in the destination volume group as a host disk (for proper isolation, an administrator should make sure that there aren’t multiple PVs on the same physical disk). When exclusive_storage is enabled in a node group, all PVs in the node group must have the same size (within a certain margin, say 1%, defined through a new parameter). Ganeti will check this condition when the exclusive_storage flag is set, whenever a new node is added and as part of cluster-verify.
When creating a new disk for an instance, Ganeti will allocate the minimum number of PVs to hold the disk, and those PVs will be excluded from the pool of available PVs for further disk creations. The underlying LV will be striped, when striping is allowed by the current configuration. Ganeti will continue to track only the LVs, and query the LVM layer to figure out which PVs are available and how much space is free. Yet, creation, disk growing, and free-space reporting will ignore any partially allocated PVs, so that PVs won’t be shared between instance disks.
For compatibility with the DRBD template and to take into account disk variability, Ganeti will always subtract 2% (this will be a parameter) from the PV space when calculating how many PVs are needed to allocate an instance and when nodes report free space.
The obvious target for this option is plain disk template, which doesn’t provide redundancy. An administrator can still provide resilience against disk failures by setting up RAID under PVs, but this is transparent to Ganeti.
When resources are dedicated and there are more spindles than instances on a node, it is natural to assign more spindles to instances than what is strictly needed. For this reason, we introduce a new resource: spindles. A spindle is a PV in LVM. The number of spindles required for a disk of an instance is specified together with the size. Specifying the number of spindles is possible only when exclusive_storage is enabled. It is an error to specify a number of spindles insufficient to contain the requested disk size.
When exclusive_storage is not enabled, spindles are not used in free space calculation, in allocation algorithms, and policies. When it’s enabled, hspace, hbal, and allocators will use spindles instead of disk size for their computation. For each node, the number of all the spindles in every LVM group is recorded, and different LVM groups are accounted separately in allocation and balancing.
There is already a concept of spindles in Ganeti. It’s not related to any actual spindle or volume count, but it’s used in spindle_use to measure the pressure of an instance on the storage system and in spindle_ratio to balance the I/O load on the nodes. These two parameters will be renamed to storage_io_use and storage_io_ratio to reflect better their meaning. When exclusive_storage is enabled, such parameters are ignored, as balancing the use of storage I/O is already addressed by the exclusive assignment of PVs.
vpcu_ratio can be used to tie the number of VCPUs to the number of CPUs provided by the hardware. We need to take into account the CPU usage of the hypervisor. For Xen, this means counting the number of VCPUs assigned to Domain-0.
For KVM, it’s more difficult to limit the number of CPUs used by the node OS. cgroups could be a solution to restrict the node OS to use some of the CPUs, leaving the other ones to instances and KVM processes. For KVM, the number of CPUs for the host system should also be a hypervisor parameter (set at the node group level).
Instances should not compete for RAM. This is easily done on Xen, but it is tricky on KVM.
Memory is already fully segregated under Xen, if sharing mechanisms (transcendent memory, auto ballooning, etc) are not in use.
Under KVM or LXC memory is fully shared between the host system and all the guests, and instances can even be swapped out by the host OS.
It’s not clear if the problem can be solved by limiting the size of the instances, so that there is plenty of room for the host OS.
We could implement segregation using cgroups to limit the memory used by the host OS. This requires finishing the implementation of the memory hypervisor status (set at the node group level) that changes how free memory is computed under KVM systems. Then we have to add a way to enforce this limit on the host system itself, rather than leaving it as a calculation tool only.
Another problem for KVM is that we need to decide about the size of the cgroup versus the size of the VM: some overhead will in particular exist, due to the fact that an instance and its encapsulating KVM process share the same space. For KVM systems the physical memory allocatable to instances should be computed by subtracting an overhead for the KVM processes, whose value can be either statically configured or set in a hypervisor status parameter.
If instances are pinned to CPUs, and the amount of memory used for every instance is proportionate to the number of VCPUs, NUMA shouldn’t be a problem, as the hypervisors allocate memory in the appropriate NUMA node. Work is in progress in Xen and the Linux kernel to always allocate memory correctly even without pinning. Therefore, we don’t need to address this problem specifically; it will be solved by future versions of the hypervisors or by implementing CPU pinning.
In order to simplify allocation and resource provisioning we want to limit the possible sizes of instances to a finite set of specifications, defined at node-group level.
Currently it’s possible to define an instance policy that limits the minimum and maximum value for CPU, memory, and disk usage (and spindles and any other resource, when implemented), independently from each other. We extend the policy by allowing it to contain more occurrences of the specifications for both the limits for the instance resources. Each specification pair (minimum and maximum) has a unique priority associated to it (or in other words, specifications are ordered), which is used by hspace (see below). The standard specification doesn’t change: there is one for the whole cluster.
For example, a policy could be set up to allow instances with this constraints: - between 1 and 2 CPUs, 2 GB of RAM, and between 10 GB and 400 GB of disk space; - 4 CPUs, 4 GB of RAM, and between 10 GB and 800 GB of disk space.
Then, an instance using 1 CPU, 2 GB of RAM and 50 GB of disk would be legal, as an instance using 4 CPUs, 4 GB of RAM, and 20 GB of disk, while an instance using 2 CPUs, 4 GB of RAM and 40 GB of disk would be illegal.
Ganeti will refuse to create (or modify) instances that violate instance policy constraints, unless the flag --ignore-ipolicy is passed.
While the changes needed to check constraint violations are straightforward, hspace behavior needs some adjustments for tiered allocation. hspace will start to allocate instances using the maximum specification with the highest priority, then it will try to lower the most constrained resources (without breaking the policy) before moving to the second highest priority, and so on.
For consistent results in capacity calculation, the specifications inside a policy should be ordered so that the biggest specifications have the highest priorities. Also, specifications should not overlap. Ganeti won’t check nor enforce such constraints, though.
We will implement this design in the following order:
In this way have always new features that are immediately useful. Spindles as a resource are not needed for correct capacity calculation, as long as allowed disk sizes are multiples of spindle size, so it’s been moved after constrained instance sizes. If it turns out that it’s easier to implement dedicated disks with spindles as a resource, then we will do that.
This section briefly describes some enhancements to the current design. They may require their own design document, and must be re-evaluated when considered for implementation, as Ganeti and the hypervisors may change substantially in the meantime.
A new resource is introduced: network bandwidth. An administrator must be able to assign some network bandwidth to the virtual interfaces of an instance, and set limits in instance policies. Also, a list of the physical network interfaces available for Ganeti use and their maximum bandwidth must be kept at node-group or node level. This information will be taken into account for allocation, balancing, and free-space calculation.
An additional enhancement is Ganeti enforcing the values set in the bandwidth resource. This can be done by configuring limits for example via openvswitch or normal QoS for bridging or routing. The bandwidth resource represents the average bandwidth usage, so a few new back-end parameters are needed to configure how to deal with bursts (they depend on the actual way used to enforce the limit).
In order to avoid unwarranted migrations between CPUs and to deal with NUMA effectively we may need CPU pinning. CPU scheduling is a complex topic and still under active development in Xen and the Linux kernel, so we wont’ try to outsmart their developers. If we need pinning it’s more to have predictable performance than to get the maximum performance (which is best done by the hypervisor), so we’ll implement a very simple algorithm that allocates CPUs when an instance is assigned to a node (either when it’s created or when it’s moved) and takes into account NUMA and maybe CPU multithreading. A more refined version might run also when an instance is deleted, but that would involve reassigning CPUs, which could be bad with NUMA.
Right now it is possible to assign more VCPUs to the instances running on a node than there are CPU available. This works as normally CPU usage on average is way below 100%. There are ways to share memory pages (e.g. KSM, transcendent memory) and disk blocks, so we could add new parameters to overcommit memory and disks, similar to vcpu_ratio.