Amazon Technologies, Inc.

Anton Valter

Alexey Gadalin

Nikolay Krasilnikov

Oleksii Tsai

Ethan John Faust

Rudresh Amin

Alexander Gregory Bestavros

Yuxuan Liu

Weili Zhong McClenahan

Diwakar Gupta

Roland Alexander Paterson-Jones

Mihir Sadruddin Surani

Scott Sikora

Guy Sebastian Parton

Andries Dippenaar

Alan Powell

Siyang Mai

Nicholas Matthew Rudi Goossens

Siwei Liu

Robert Beazley

Mykhailo Karataiev

Saurabh Nirmal Modh

Techniques for managing cloud computing resources hosting burstable performance instances are described. A host computer system of a provider network executes burstable performance compute instances. Compute capacity usage data is obtained from the host computer system, the compute capacity usage data including a first indication of a first compute capacity used by a first burstable performance compute instance. A first weight for the first burstable performance compute instance is calculated, the first weight being inversely related to the first compute capacity. A scheduler of the host computer system is updated with process prioritization weights, the process prioritization weights including a first process prioritization weight that is based at least in part on the first weight. The scheduler allocates compute capacity based on the process prioritization weights.

Managing fairness of resource scheduling for performance bursting in multi-tenant environments

This disclosure describes techniques for increasing the baseline performance of a burstable instance to an increased performance level for a limited time period. For example, a user may schedule a time period which the burstable instance has access to 100% of a CPU, instead of competing with other burstable instances even during periods of bursting as in prior techniques. In some cases, the user uses credits to reserve the time periods at which the increased performance level is requested. After receiving the reservation, a resource system selects computing resources to host the burstable instance such that the burstable instance, can operate at the requested increased performance level. After the time period has ended, the burstable instance may return to the baseline performance level.

Reserving increased performance of burstable instances for a limited time period

Techniques and technologies for migrating workloads between different VM instance types based on changes in resource utilization over a period of time. Selecting one VM instance type to host workloads can result in use of a VM instance type that is optimized for the workload during one timeframe of a period of time, but is over-provisioned or under-provisioned for the workload during another timeframe of a period of time. Accordingly, a service provider network may analyze historical utilization patterns of workloads over a period of time, and identify different VM instance types that are optimized for the workloads during different timeframes of the period of time. The service provider network may then cause the workloads to be migrated between the different VM instance types over the period of time such that the workloads are hosted on VM instances that are optimized for the current resource utilization of the workloads.

Migrating workloads between computing platforms according to resource utilization

Generally described, one or more aspects of the present application relate to an instance cloning service for cloning a running virtual machine instance and configuring the resources of the new virtual machine instance created by cloning the running virtual machine instance. For example, the instance cloning service can create the new virtual machine instance and attach it to a storage volume to which the original virtual machine instance is attached. The instance cloning service may subsequently move the new virtual machine instance to another storage volume (e.g., based on a user request or based on the performance of the instances). Alternatively, upon cloning, the instance cloning service may attach the new virtual machine instance to a new storage volume. For example, the instance cloning service may use copy-on-write to provide the new virtual machine instance with its own storage volume.

Virtual machine cloning and resource configuration

Disclosed are various embodiments for an emulation layer for overcoming incompatibilities in virtual machine instance migration. In one embodiment, a virtual machine instance configured for a first computing architecture is executed in a first computing device having the first computing architecture. The virtual machine instance from the first computing device is migrated to a second computing device having a second computing architecture. The virtual machine instance is executed in the second computing device using an emulation layer that emulates the first computing architecture in the second computing architecture.

Emulation layer for overcoming incompatibilities in virtual machine instance migration

This disclosure leverages multi-attach to block store volumes for more reliable live migration of virtualized resources. A block storage client of a virtualized resource operating on a source host in a first data center can be attached to a block storage volume stored on block storage hosts in the first data center. State data associated with the virtual machine can be transmitted from the source host to a target host, after which the virtual machine can run on the target host and operations of the virtualized resources may be ceased on the source host. Failure of the migration may require roll back to the source host. The source host may remain connected to the volume while the target host client connects to the volume, such that the volume may be accessed by the block storage client on the source host after rollback to provide uninterrupted operation of the virtual machine.

Multiple simultaneous volume attachments for live migration between cloud regions and edge locations

Techniques for a device with NUMA a memory architecture to migrate virtual resources between NUMA nodes to reduce resource contention between virtual resources running on the NUMA nodes. In some examples, the device monitors various metrics and/or operations of the NUMA nodes and/or virtual resources, and detect events that indicate that virtual resources running on a same NUMA node are contending, or are likely to contend, for computing resources of the NUMA node. Upon detecting such an event, the device may migrate a virtual resource from the NUMA node to another NUMA node on the device that has an availability of computing resources. The device may then migrate the virtual resource from the overcommitted NUMA node onto the NUMA node that has availability to run the virtual resource. In this way, devices may reduce resource contention among virtual resources running on a same NUMA node.

Migrating virtual resources between non-uniform memory access (NUMA) nodes

This disclosure describes systems, devices, and techniques for migrating virtualized resources between the main region and edge locations. Live migration enables virtualized resources to remain operational during migration. Edge locations are typically separated from secure data centers via the Internet, a direct connection, or some other intermediate network. Accordingly, to place virtualized resources within an edge location, the virtualized resources must be migrated over a secure communication tunnel that can protect virtualized resource data during transmission over the intermediate network. The secure communication tunnel may have limited data throughput. To efficiently utilize resources of the secure communication tunnel, and to reduce the impact of migrations on virtualized resource operations, virtualized resource migrations may be carefully scheduled in advance. For instance, virtualized resources may be selectively migrated at times-of-day in which they are likely to be relatively idle, or at times when the communication tunnel is predicted to have sufficient bandwidth.

Dynamic scheduling for live migration between cloud regions and edge locations

This disclosure describes systems, devices, and techniques for live migrating virtualized resources between the main region and edge locations. Live migration enables virtualized resources to remain operational during migration. Edge locations are typically separated from secure data centers via the Internet, a direct connection, or some other intermediate network. Accordingly, to place virtualized resources within an edge location, the virtualized resources must be migrated over a secure communication tunnel that can protect virtualized resource data during transmission over the intermediate network. The secure communication tunnel may have limited data throughput. To efficiently utilize resources of the secure communication tunnel, virtualized resource data may be transferred over the tunnel in a two-stage process. The disclosed technology provides various techniques for dynamically selecting which virtualized resource data, and how much virtualized resource data, is transferred over each of the pre-copy and post-copy stages.

Dynamic pre-copy and post-copy determination for live migration between cloud regions and edge locations

This disclosure describes systems, devices, and techniques for quick restarts of virtualized resources hosted by a cloud-based network. In an example method, a request to restart a virtualized resource hosted by a first server can be received. A second server, with a capacity to host the virtualized resource, may be identified. In response to identifying the second server, the virtualized resource can be migrated from the first server to the second server. A message confirming that the virtualized resource has been restarted may be transmitted.

Quick restart using virtualized resource migration

Methods and apparatus for post data synchronization in live migration of domains. Host devices on a network may implement virtual machines (VMs) as domains in an execution environment, and may provide local persistent storage for data of the VMs. A migration technique for moving a domain including the persistent data from one host device to another host device is described in which the VM is instantiated on the target device, and the domain is switched to the target device. Synchronization of the VM's data from the persistent storage on the source device to the target device is then initiated, for example according to a distributed replicated storage technique that makes the target device's persistent storage the primary storage and the source device's persistent storage the secondary storage for the VM. Once the data is synchronized, the VM and its respective storage on the source device are released.

Post data synchronization for domain migration

Techniques for managing large-scale automatic fleet refresh operations are described herein. An application programming interface request to perform a refresh operation on a set of computer system instances is received. The application programming interface request includes a set of constraints for performing the refresh operation which are used to determine the impact of performing the refresh operation. Based at least in part on the impact, a set of schedules for performing the refresh operation is provided.

Growing community of inventors

Bellevue, WA, United States of America

Anton Valter