Patent Application
20250219942
A data center houses computer systems and various networking, storage, and other related components. Data centers, for example, are used by service providers to provide computing services to businesses and individuals as a remote computing service or provide “software as a service” (e.g., cloud computing). Software defined networking (SDN) enables centralized configuration and management of physical and virtual network devices as well as dynamic and scalable implementation of network policies. Network bandwidth and latency are critical constraints for some workloads. For example, for artificial intelligence (AI) workloads, large amounts of data (e.g., weights) need to be moved between successive AI training/inferencing phases. The efficient processing of workloads and efficiently utilizing the physical and virtual network devices are important for maintaining scalability and efficient operation in such networks.
It is with respect to these considerations and others that the disclosure made herein is presented.
The present disclosure describes an enhanced router/smart switch that improves network performance (latency and bandwidth) for workloads by providing in-network primitives that improve the performance of operations such as Broadcast and Gather and Reduce that are used, for example, in AI workloads. The enhanced router/smart switch leverages the observation that the Broadcast and Reduce primitives are more performant (e.g., reduced latency and bandwidth) when performed in the network rather than in graphics processing unites (GPUs) or central processing units (CPUs) which are often deployed as leaf nodes in a typical network topology in a datacenter. Broadcast and Reduce primitives are more performant when performed in the network because in-network processing avoids duplicate transmission of identical data and reduces latency by utilizing the inherent tree topology in the network rather than building a tree topology among the GPUs or CPUs.
In an SDN that implements an enhanced router/smart switch for processing specified workloads, primitives are implemented in the network processors in the enhanced router. The primitives operate on communication traffic that is specified in collectives: all-gather, all-reduce, reduce-scatter, etc. These primitives reduce the time to finish collectives (e.g., lower latency) or reduce the amount of network traffic from collectives (e.g., lower total bandwidth usage), allowing for faster training, inference, and more efficient servicing of workloads. Examples included in this disclosure include multicast, in-network aggregation, and gradient compression.
The described techniques can allow for virtual computing environments to support a variety of workload configurations while maintaining efficient use of computing resources such as processor cycles, memory, network bandwidth, and power. This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended that this Summary be used to limit the scope of the claimed subject matter. Furthermore, the claimed subject matter is not limited to implementations that solve any or all disadvantages noted in any part of this disclosure.
The Detailed Description is described with reference to the accompanying figures. In the description detailed herein, references are made to the accompanying drawings that form a part hereof, and that show, by way of illustration, specific embodiments or examples. The drawings herein are not drawn to scale. Like numerals represent like elements throughout the several figures.
Network bandwidth and latency are critical constraints for certain workloads. For example, in AI workloads, large amounts of data (e.g., weights) need to be moved between successive AI training/inferencing phases. Network performance is thus a key determinant for the completion time of a workload. Additionally, by enabling fast movement of data, the capabilities of the network can reduce the time to completion and improve GPU efficiency by driving up network utilization. While the following description describes embodiments in the AI workload context, it should be understood that the disclosed principles are applicable to other types of workloads.
The disclosed embodiments enable datacenters to provide services in a manner that can enhance system flexibility and efficiency while reducing cost and complexity, allowing for more efficient use of computing, storage, and network resources. Efficient implementation of the end-to-end services by a cloud service provider can enable an experience that is seamless and more consistent across various footprints. The effective distribution of the described disaggregation and pooling techniques can also be determined based on the implications for various performance and security implications such as latency and data security.
The various embodiments disclosed herein provide an enhanced router/smart switch that improves network performance, for example, for AI workloads by providing in-network primitives that improve the performance of Broadcast and Gather and Reduce operations used in AI workloads. As used herein, the enhanced router may also be referred to as a smart switch or Smartswitch and can be referred to as an AI router when used in the context of AI workloads.
In the example of AI workloads, there are several primitive AI workload operations that drive network traffic, such as:
The disclosed embodiments leverage the observation that primitives such as the Broadcast and Reduce primitives are more performant (e.g., reduced latency and bandwidth) when performed in the network rather than in GPUs/CPUs which are deployed as leaf nodes in a typical network topology in a datacenter. Broadcast and Reduce primitives are more performant when performed in the network because in-network processing avoids duplicate transmission of identical data and reduces latency by utilizing the inherent topology of the network rather than building a tree topology among the GPUs/CPUs.
In an SDN that implements an enhanced router for processing workloads such as AI workloads, primitives are implemented in the network processors in the enhanced router. The primitives operate on communication traffic that is specified in collectives: all-gather, all-reduce, reduce-scatter, etc. These primitives reduce the time to finish collectives (e.g., lower latency) or reduce the amount of network traffic from collectives (e.g., lower total bandwidth usage), allowing for faster training, inference, and more efficient servicing of ML workloads. Examples included in this disclosure include multicast, in-network aggregation, and gradient compression.
In an embodiment, an application programming interface (API) is implemented to provide a standard interface to abstract underlying hardware differences and allow interoperability between enhanced router devices. The API allows for one-to-one transformations, one-to-many transformations, and many-to-one transformations when implementing the primitives.
The enhanced router can be an Ethernet switch enhanced with data processing units (DPUs) and configured to provide:
Additionally, the enhanced router can be deployed as a Tier 0 or a T1 switch or multiple devices can be deployed at various places in the network, allowing for the enhanced routers to be incrementally used to enhance network performance by offloading work from CPUs/GPUs.
To enable broadcast or multicast support, each workload may be identified using a unique ID such as a VXLAN ID. The described examples are illustrated using VXLAN, but it should be understood that the disclosed principles can be implemented using other tunneling technologies. If the destination address of a packet in the VXLAN header indicates a broadcast IP address or the address of the multicast group, the enhanced router forwards that packet to all nodes in the associated group based on the VNET ID or multicast group address. The enhanced router is programmed with group membership information by a network controller using the extension APIs.
To enable Reduce (Summation) support, the enhanced router provides Summation support for multiple summation operations that are in progress and support for retransmission of same data. The enhanced router implements a mechanism for eliminating duplicate data. The enhanced router further implements a partial computation where a portion of the computation can be completed at the GPU/CPU, thereby enabling coexistence with legacy mechanisms to compute at the GPU/CPU. Each packet with data that needs to be aggregated identifies the aggregation operation by indicating the operation ID and operation type as well as the data size and location within the packet. This information can be included in VXLAN extension headers, or other types of headers in various tunneling protocols. The packets also specify other salient information regarding the aggregation such as the output format and the aggregation delay period. The results of the aggregation operation can be sent to one or all members of a specified workload group.
Referring to the appended drawings, in which like numerals represent like elements throughout the several FIGURES, aspects of various technologies for AI support will be described. In the following detailed description, references are made to the accompanying drawings that form a part hereof, and which are shown by way of illustration specific configurations or examples.
Referring to
The Smartswitch 110 receives an input data packet 122, for example via cloud 105. The input data packet 122 can be addressed to an endpoint hosted by a VM 131 of the user network 134. The Smartswitch 110 can facilitate operations such as Broadcast 116, Aggregate 117, as well as other functions 118. The Smartswitch 110 forwards the input data packet 122 to a hardware-based network interface device such as NIC 133. In some embodiments, the forwarding of the input data packet 122 can be facilitated by fabric 102 that can include various networking devices such as switches.
In an example, an AI workload is identified as being processed in the computing network 100. A destination address 123 of a header 121 of the data packet 122 is determined. In response to determining that the destination address 123 indicates a broadcast IP address, the data packet 122 broadcast to all nodes associated with an ID such as a VNET ID 135 of the AI workload. An aggregation operation of the data packet 122 is identified based on an operation ID, operation type, and data ID in an extension header (such as a tunneling header) 124 of the data packet 122. In response to determining that the data packet 122 needs to be aggregated, results of the aggregation operation are broadcast to all nodes associated with the VNET ID 135 of the AI workload. Operations can be performed on data 125 of the packet 122.
The present disclosure includes network processor primitives to assist machine learning workloads. In an embodiment, the primitives can be implemented in network interface cards (NICs) that are directly attached to servers, in switches and routers in the network infrastructure, disaggregated pools of NICs, and the like.
At least some of the network processor primitives include those related to communication traffic that is specified in collectives, for example all-gather, all-reduce, reduce-scatter, and the like. Reducing the time to finish collectives (e.g., lower latency) or reducing the amount of network traffic from collectives (e.g., lower total bandwidth usage) enables faster training, inference, and lower implementation costs to service ML workloads.
One example of a primitive that the described API enables offloading to network processors includes multicast. Many collectives, including all-gather and reduce-scatter, send the same data to multiple recipients. Using multicast to carry such data reduces the number of copies that are carried on the network.
One example of a primitive that the described API enables for offloading to network processors includes in-network aggregation. Some collectives such as all-reduce perform associative and commutative operations on groups of data items. For example, all-reduce is commonly used to compute gradients, where each participating GPU sends its partial gradient to a recipient which computes a sum of the gradients. The sum must be broadcast to all participating GPUs which would benefit from the above multicast primitive.
With in-network aggregation, the aggregation operation is performed in the network processors. That is, a switch which receives a few of the partial gradients can emit the sum of these gradients. In-network aggregation reduces the amount of data carried on the network as well as amount of the computation at the host GPUS (to aggregate) and speeds up the computation by reducing communication time for data.
One example of a primitive that the described API enables offloading to network processors includes gradient compression. Gradients can be compressed to trade-off reduction in network communication costs (latency and bandwidth) for answer quality (e.g., approximate convergence of training models). Examples include reducing the size of floating points, applying learned (e.g., deep neural network) transformations on the gradients, and the like.
In an embodiment, an application programming interface (API) is provided to allow multiple entities to provide implementations of the primitives while allowing for interoperability. Additionally, the on-processor implementations can be decoupled from the control plane that will use these implementations. Thus, the primitives are only specified at the granularity of what each processor performs. The control path specifies the operations at each processor (e.g., which packet fields to read or write into) so that interoperability is enabled. Additionally, the API abstracts the network technology (e.g., Infiniband). The API is operable to support a wide range of primitives including but not limited to those described herein. In various embodiments, the API is implemented in network processors.
In an embodiment, the API supports transformations for at least three classes:
One-to-one transformations: These transformations take as input a single packet and emit a packet. The API allows for a broad class of per-packet transformations if the transformations are effectively stateless, so that they can also be implemented in parallel without requesting support for coherence or memory. Examples of such transformations include degrading floating point resolution and gradient compression.
One-to-many transformations: These transformations can emit multiple packets for one packet. Multicast is an example of this transformation. The API allows for a broad class of one-to-many transformations ranging from flexibility in (a) which packets to transform and (b) the kinds of changes to make in the packets that will be output. In an embodiment, the transformations are effectively stateless.
Many-to-one transformations: These transformations specify grouping criteria on input packets and emit one output packet per group. For example, in the case of in-network aggregation, the grouping criteria may be the epoch identifier and the output can be a sum of the gradients received that have the same epoch identifier. In an embodiment, the API specifies (a) a matching clause to identify a group, (b) the time-duration for how long to wait before emitting, and (c) the transformation on the packet which includes the commutative and associative operation to perform on packet contents (e.g., sum).
These transformations are stateful. To assess the amount of state per currently active group, the memory footprint may be as small as one packet. That is, it may not be necessary to store all packets that match in the group. For example, when computing gradient sum, the memory footprint is one packet—the sum (or average) of all the gradients received thus far.
Some classes of transformations can overlap with or be used in conjunction with transformations from other classes. For example, a one-to-many transformation such as multicast may also send compressed gradients. Furthermore, the classes can be combined to obtain new net transformations such as many-to-many transformations where the output of a group's sum is sent to multiple recipients.
The primitives supported by the API are applicable to applications beyond machine learning (ML) workloads. For example, virtual-taps, wherein packets matching a condition (or all packets) are forwarded to a separate diagnostic service point can be supported in this API as a one-to-many transformation. That is, packets matching the tapping criteria will be forwarded not only to their original destination but also to the diagnostic service point. As another example, aspects of INT telemetry such as counting different aspects (numbers of bytes, packets that match a criteria) can be implemented using many-to-one transformations in this API. NDP, which removes the data body of a packet and forwards just the packet headers onwards instead of dropping packets when queues start building up, can also be implemented as a one-to-one transformation. The ability to change parts of a packet header based on a match condition can also be used to enable many different routing and load balancing schemes.
For the control path, embodiments may include simple common constructs such as:
For the data path, matching functions will identify which packets to apply a transform upon. The matching functions may be supported across all transformations and comprise the offset from which to pick up the values of the identifier and the Boolean valued filter expressions.
Boolean operations can be equality, less than, greater than, and the like. An example matching function is ‘destination IP address matches a multicast group value.’
For transformation specifiers:
The disclosure provides a protocol that declares transformations that can be executed by NICs, switches, or appliances within the network. In an embodiment, the protocol may be referred to as In-network Transformation Protocol (INTP). The protocol enables interoperability across multiple implementors and decouples changes that must be made to the host driver from changes in the network elements. In an embodiment, the header can be included in RDMA over Converged Ethernet or Infiniband or other transport mechanism.
In an embodiment, message-sized headers (e.g., using jumbo frames) can be implemented.
In the example of multicast, the source requests multicast transformations and specifies the corresponding group-address. The group address may be the list of GPUs participating in an all-gather collective.
The control plane protocol is responsible for communicating with appropriate network elements on the path that they must honor this transformation. The mapping between group addresses to the list of individual addresses that belong to the group can be performed using a standard protocol such as IGMP or have this mapping specified by the control plane.
The source may send such packets towards a network element that honors the transformation. More implicitly, a ToR switch on the path may implement the transformation. A network element that processes this packet edits the outer packet header with the IP addresses of the individual constituents of the group.
As shown in
In the example of gradient compression, floating point, and the like, techniques can be implemented that reduce message sizes and floating-point representations to reduce the amount of computation involved. Some examples include MSFP, TensorRT, and FlexPoint.
A network element that honors this transformation will store the tensor for some amount of time (either implicitly determined or specified by the control plane protocol). All tensors obtained that have the same epoch ID will be coalesced into a single packet using the aggregation operation and the list of aggregatees will be expanded.
As with the case of multicast above, in-network aggregation can be applied recursively by more than one network element on a given network path. Additionally, there may not be any network elements on a network path that implement network aggregation. In such cases, the receiver will obtain the tensor data from individual sources.
At the source, a driver is configured to add the appropriate transformation header. Sources may also change what they emit. For example, with multicast, the sources emit one packet to the group address and not one packet per group constituent.
At the receiver, the driver is configured to be aware of and consume the transformation header. Receivers may also change how they handle, for example, in-network aggregation. Receivers should expect to receive fewer packets than one packet per group constituent.
Determining which network element(s) honor which transformation is the responsibility of the control protocol. The transformations are singletons, i.e., specific to individual network elements.
Similar to SDN goal-state dissemination: each network element that participates in the transformation may subscribe to an item and the controller pushes changes. Other implementations are possible as well such as those relying on OpenFlow.
For certain transformations, the end-hosts (e.g., a NCCL driver) are configured to introspect the availability of corresponding transformation along their network paths. For example, with multicast, a source knows whether to assume that multicast will be supported.
As another example, for collective scheduling, to overlap computation and communication, an end-host is configured to know the effective topology for the collective. In various embodiments, multiple implementors of these transformations can interoperate with each other.
Data center 200 may correspond to network 100 in
Referring to
Communications network 220 may provide access to computers 202. Computers 202 may be computers utilized by users 201. Computer 202a, 202b or 202c may be a server, a desktop or laptop personal computer, a tablet computer, a smartphone, a set-top box, or any other computing device capable of accessing data center 200. User computer 202a or 202b may connect directly to the Internet (e.g., via a cable modem). User computer 202c may be internal to the data center 200 and may connect directly to the resources in the data center 200 via internal networks. Although only three user computers 202a,202b, and 202c are depicted, it should be appreciated that there may be multiple user computers.
Computers 202 may also be utilized to configure aspects of the computing resources provided by data center 200. For example, data center 200 may provide a Web interface through which aspects of its operation may be configured through the use of a Web browser application program executing on user computer 202. Alternatively, a stand-alone application program executing on user computer 202 may be used to access an application programming interface (API) exposed by data center 200 for performing the configuration operations.
Servers 226 may be configured to provide the computing resources described above. One or more of the servers 226 may be configured to execute a manager 220a or 220b (which may be referred herein singularly as “a manager 230” or in the plural as “the managers 230”) configured to execute the virtual machines. The managers 230 may be a virtual machine monitor (VMM), fabric controller, or another type of program configured to enable the execution of virtual machines 222 on servers 226, for example.
It should be appreciated that although the embodiments disclosed above are discussed in the context of virtual machines, other types of implementations can be utilized with the concepts and technologies disclosed herein.
In the example data center 200 shown in
It should be appreciated that the network topology illustrated in
It should also be appreciated that data center 200 described in
In some embodiments, aspects of the present disclosure may be implemented in a mobile edge computing (MEC) environment implemented in conjunction with a 4G, 5G, or other cellular network. MEC is a type of edge computing that uses cellular networks and 5G and enables a data center to extend cloud services to local deployments using a distributed architecture that provide federated options for local and remote data and control management. MEC architectures may be implemented at cellular base stations or other edge nodes and enable operators to host content closer to the edge of the network, delivering high-bandwidth, low-latency applications to end users. For example, the cloud provider's footprint may be co-located at a carrier site (e.g., carrier data center), allowing for the edge infrastructure and applications to run closer to the end user via the 5G network.
In some of the illustrated example scenarios described herein, SDN capabilities may be enhanced by disaggregating policy enforcement from the host and moving it elsewhere on the network, such as onto an SDN appliance. Software defined networking (SDN) is conventionally implemented on a general-purpose compute node. The SDN control plane may program the host to provide core network functions such as security, virtual network, and load balancer policies. An SDN appliance can be used to host these agents and provide switch functionality, and can further provide transformations and connectivity. The SDN appliance can accept policies that perform transformations. In some embodiments, an agent can be implemented that programs the drivers that run on the SDN appliance. The traffic sent by workloads can be directed through the SDN appliance, which can apply policies and perform transformations on the traffic and send the traffic to the destination. In some configurations, the SDN appliance may include a virtual switch such as a virtual filtering platform.
It should be appreciated that the subject matter presented herein may be implemented as a computer process, a computer-controlled apparatus, a computing system, an article of manufacture, such as a computer-readable storage medium, or a component including hardware logic for implementing functions, such as a field-programmable gate array (FPGA) device, a massively parallel processor array (MPPA) device, a graphics processing unit (GPU), an application-specific integrated circuit (ASIC), a multiprocessor System-on-Chip (MPSoC), etc.
A component may also encompass other ways of leveraging a device to perform a function, such as, for example, a) a case in which at least some tasks are implemented in hard ASIC logic or the like; b) a case in which at least some tasks are implemented in soft (configurable) FPGA logic or the like; c) a case in which at least some tasks run as software on FPGA software processor overlays or the like; d) a case in which at least some tasks run as software on hard ASIC processors or the like, etc., or any combination thereof. A component may represent a homogeneous collection of hardware acceleration devices, such as, for example, FPGA devices. On the other hand, a component may represent a heterogeneous collection of different types of hardware acceleration devices including different types of FPGA devices having different respective processing capabilities and architectures, a mixture of FPGA devices and other types hardware acceleration devices, etc.
Turning now to
It should also be understood that the illustrated methods can end at any time and need not be performed in their entireties. Some or all operations of the methods, and/or substantially equivalent operations, can be performed by execution of computer-readable instructions included on a computer-storage media, as defined herein. The term “computer-readable instructions,” and variants thereof, as used in the description and claims, is used expansively herein to include routines, applications, application modules, program modules, programs, components, data structures, algorithms, and the like. Computer-readable instructions can be implemented on various system configurations, including single-processor or multiprocessor systems, minicomputers, mainframe computers, personal computers, hand-held computing devices, microprocessor-based, programmable consumer electronics, combinations thereof, and the like.
It should be appreciated that the logical operations described herein are implemented (1) as a sequence of computer implemented acts or program modules running on a computing system such as those described herein) and/or (2) as interconnected machine logic circuits or circuit modules within the computing system. The implementation is a matter of choice dependent on the performance and other requirements of the computing system. Accordingly, the logical operations may be implemented in software, in firmware, in special purpose digital logic, and any combination thereof. Thus, although the routine 300 is described as running on a system, it can be appreciated that the routine 300 and other operations described herein can be executed on an individual computing device or several devices.
Referring to
Operation 303 illustrates determining, by the network device, that the packet is associated with the workload being processed in the computing network.
Operation 305 illustrates based on information contained in a header of the packet, performing, by the network device, an operation on a payload of the packet.
Operation 307 illustrates sending, by the network device, the processed packet to a destination address determined based on the header of the packet or additional configuration information.
Turning now to
Referring to
Operation 353 illustrates receiving, by the network device, a data packet associated with the workload being processed in the computing network.
Operation 355 illustrates determining, by the network device, that the packet is associated with the workload being processed in the computing network based on a transmission identifier and a parameter indicated by the primitive. In an embodiment, the transmission identifier and parameter are contained in a header of the data packet.
Operation 357 illustrates in response to the determining, performing, by the network device, the operation on a payload of the data packet.
Operation 359 illustrates sending, by the network device, the processed data packet to a destination address identified in the header of the packet.
In various embodiments, computing device 400 may be a uniprocessor system including one processor 440 or a multiprocessor system including several processors 44 (e.g., two, four, eight, or another suitable number). Processors 44 may be any suitable processors capable of executing instructions. For example, in various embodiments, processors 44 may be general-purpose or embedded processors implementing any of a variety of instruction set architectures (ISAs), such as the x44, PowerPC, SPARC, or MIPS ISAs, or any other suitable ISA. In multiprocessor systems, each of processors 44 may commonly, but not necessarily, implement the same ISA.
System memory 420 may be configured to store instructions and data accessible by processor(s) 440. In various embodiments, system memory 420 may be implemented using any suitable memory technology, such as static random access memory (SRAM), synchronous dynamic RAM (SDRAM), nonvolatile/Flash-type memory, or any other type of memory. In the illustrated embodiment, program instructions and data implementing one or more desired functions, such as those methods, techniques and data described above, are shown stored within system memory 420 as code 425 and data 424.
In one embodiment, I/O interface 430 may be configured to coordinate I/O traffic between the processor 440, system memory 420, and any peripheral devices in the device, including network interface 410 or other peripheral interfaces. In some embodiments, I/O interface 430 may perform any necessary protocol, timing, or other data transformations to convert data signals from one component (e.g., system memory 420) into a format suitable for use by another component (e.g., processor 440). In some embodiments, I/O interface 430 may include support for devices attached through various types of peripheral buses, such as a variant of the Peripheral Component Interconnect (PCI) bus standard or the Universal Serial Bus (USB) standard, for example. In some embodiments, the function of I/O interface 430 may be split into two or more separate components. Also, in some embodiments some or all of the functionality of I/O interface 430, such as an interface to system memory 420, may be incorporated directly into processor 440.
Network interface 410 may be configured to allow data to be exchanged between computing device 400 and other device or devices 460 attached to a network or network(s) 450, such as other computer systems or devices as illustrated in
In some embodiments, system memory 420 may be one embodiment of a computer-accessible medium configured to store program instructions and data as described above for the Figures for implementing embodiments of the corresponding methods and apparatus. However, in other embodiments, program instructions and/or data may be received, sent or stored upon different types of computer-accessible media. A computer-accessible medium may include non-transitory storage media or memory media, such as magnetic or optical media, e.g., disk or DVD/CD coupled to computing device 400 via I/O interface 430. A non-transitory computer-accessible storage medium may also include any volatile or non-volatile media, such as RAM (e.g. SDRAM, DDR SDRAM, RDRAM, SRAM, etc.), ROM, etc., that may be included in some embodiments of computing device 400 as system memory 420 or another type of memory. Further, a computer-accessible medium may include transmission media or signals such as electrical, electromagnetic or digital signals, conveyed via a communication medium such as a network and/or a wireless link, such as may be implemented via network interface 410. Portions or all of multiple computing devices, such as those illustrated in
Various storage devices and their associated computer-readable media provide non-volatile storage for the computing devices described herein. Computer-readable media as discussed herein may refer to a mass storage device, such as a solid-state drive, a hard disk or CD-ROM drive. However, it should be appreciated by those skilled in the art that computer-readable media can be any available computer storage media that can be accessed by a computing device.
By way of example, and not limitation, computer storage media may include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules or other data. For example, computer media includes, but is not limited to, RAM, ROM, EPROM, EEPROM, flash memory or other solid state memory technology, CD-ROM, digital versatile disks (“DVD”), HD-DVD, BLU-RAY, or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by the computing devices discussed herein. For purposes of the claims, the phrase “computer storage medium,” “computer-readable storage medium” and variations thereof, does not include waves, signals, and/or other transitory and/or intangible communication media, per se.
Encoding the software modules presented herein also may transform the physical structure of the computer-readable media presented herein. The specific transformation of physical structure may depend on various factors, in different implementations of this description. Examples of such factors may include, but are not limited to, the technology used to implement the computer-readable media, whether the computer-readable media is characterized as primary or secondary storage, and the like. For example, if the computer-readable media is implemented as semiconductor-based memory, the software disclosed herein may be encoded on the computer-readable media by transforming the physical state of the semiconductor memory. For example, the software may transform the state of transistors, capacitors, or other discrete circuit elements constituting the semiconductor memory. The software also may transform the physical state of such components in order to store data thereupon.
As another example, the computer-readable media disclosed herein may be implemented using magnetic or optical technology. In such implementations, the software presented herein may transform the physical state of magnetic or optical media, when the software is encoded therein. These transformations may include altering the magnetic characteristics of particular locations within given magnetic media. These transformations also may include altering the physical features or characteristics of particular locations within given optical media, to change the optical characteristics of those locations. Other transformations of physical media are possible without departing from the scope and spirit of the present description, with the foregoing examples provided only to facilitate this discussion.
In light of the above, it should be appreciated that many types of physical transformations take place in the disclosed computing devices in order to store and execute the software components and/or functionality presented herein. It is also contemplated that the disclosed computing devices may not include all of the illustrated components shown in
Although the various configurations have been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended representations is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as example forms of implementing the claimed subject matter.
Conditional language used herein, such as, among others, “can,” “could,” “might,” “may,” “e.g.,” and the like, unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey that certain embodiments include, while other embodiments do not include, certain features, elements, and/or steps. Thus, such conditional language is not generally intended to imply that features, elements, and/or steps are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without author input or prompting, whether these features, elements, and/or steps are included or are to be performed in any particular embodiment. The terms “comprising,” “including,” “having,” and the like are synonymous and are used inclusively, in an open-ended fashion, and do not exclude additional elements, features, acts, operations, and so forth. Also, the term “or” is used in its inclusive sense (and not in its exclusive sense) so that when used, for example, to connect a list of elements, the term “or” means one, some, or all of the elements in the list.
While certain example embodiments have been described, these embodiments have been presented by way of example only, and are not intended to limit the scope of the inventions disclosed herein. Thus, nothing in the foregoing description is intended to imply that any particular feature, characteristic, step, module, or block is necessary or indispensable. Indeed, the novel methods and systems described herein may be embodied in a variety of other forms; furthermore, various omissions, substitutions and changes in the form of the methods and systems described herein may be made without departing from the spirit of the inventions disclosed herein. The accompanying claims and their equivalents are intended to cover such forms or modifications as would fall within the scope and spirit of certain of the inventions disclosed herein.
It should be appreciated any reference to “first,” “second,” etc. items and/or abstract concepts within the description is not intended to and should not be construed to necessarily correspond to any reference of “first,” “second,” etc. elements of the claims. In particular, within this Summary and/or the following Detailed Description, items and/or abstract concepts such as, for example, individual computing devices and/or operational states of the computing cluster may be distinguished by numerical designations without such designations corresponding to the claims or even other paragraphs of the Summary and/or Detailed Description. For example, any designation of a “first operational state” and “second operational state” of the computing cluster within a paragraph of this disclosure is used solely to distinguish two different operational states of the computing cluster within that specific paragraph—not any other paragraph and particularly not the claims.
Although the various techniques have been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended representations is not necessarily limited to the specific features or acts described. Rather, the specific features and acts are disclosed as example forms of implementing the claimed subject matter.
The disclosure presented herein also encompasses the subject matter set forth in the following clauses:
Clause 1: A method for processing data packets in a computing network processing workloads in a virtualized computing network comprising a plurality of computing nodes and network devices configured to implement a software defined network (SDN), the computing nodes hosting virtual machines or containers and the network devices configured to process data packets associated with the workloads outside of the computing nodes, the method comprising:
Clause 2: The method of clause 1, wherein the workload comprises an artificial intelligence (AI) workload.
Clause 3: The method of any of clauses 1-2, wherein determining the destination address comprises determining a VXLAN header of the data packet.
Clause 4: The method of any of clauses 1-3, further comprising:
Clause 5: The method of any of clauses 1-4, further comprising:
Clause 6: The method of any of clauses 1-5, further comprising:
Clause 7: The method of any of clauses 1-6, wherein the workload is identified with a unique VXLAN ID.
Clause 8: The method of any of clauses 1-7, wherein the network device is a switch configured with data processing units (DPUs).
Clause 9: The method of any of clauses 1-8, wherein the switch is deployed as a Tier 0 switch.
Clause 10: A network device configured to manage data packets in a computing network processing artificial intelligence (AI) workloads, the network device configured to perform operations comprising:
Clause 11: The network device of clause 10, wherein determining the destination address comprises determining a VXLAN header of the data packet and wherein the network device is configured to perform operations comprising:
Clause 12: The network device of any of clauses 10 and 11, wherein the network device is configured to perform operations comprising:
Clause 13: The network device of any of clauses 10-12, wherein the network device is configured to perform operations comprising:
Clause 14: The network device of any of clauses 10-13, wherein the AI workload is identified with a unique VXLAN ID.
Clause 15: The network device of any of clauses 10-14, wherein the network device is a switch configured with data processing units (DPUs).
Clause 16: The network device of any of clauses 10-15, wherein the switch is deployed as a Tier 0 switch.
Clause 17: A system for processing data packets in a computing network processing artificial intelligence (AI) workloads in a virtualized computing network comprising a plurality of computing nodes and network devices configured to implement a software defined network (SDN), the computing nodes hosting virtual machines or containers and the network devices configured to process data packets associated with the AI workloads outside of the computing nodes, the network device configured to perform operations comprising:
Clause 18: The system of clause 17, wherein determining the destination address comprises determining a VXLAN header of the data packet.
Clause 19: The system of any of clauses 17 and 18, wherein the network device is a switch configured with data processing units (DPUs).
Clause 20: The system of any of the clauses 17-19, wherein the AI workload is identified with a unique VXLAN ID.
The disclosure presented herein also encompasses the subject matter set forth in the following additional set of clauses:
Clause 1: A method for processing data packets in a computing network processing workloads in a network device operating in a software defined network (SDN), the network device configured to execute primitives for data payloads in the packets associated with the workloads in the network prior to forwarding the packets to hosts in the SDN, the method comprising:
Clause 2: The method of clause 1, wherein the workload comprises an artificial intelligence (AI) workload.
Clause 3: The method of any of clauses 1-2, wherein the network device comprises one of a network interface card (NIC), a network switch, or a disaggregated pool of NICs.
Clause 4: The method of any of clauses 1-3, wherein the workload comprises a collective.
Clause 5: The method of any of clauses 1-4, wherein the operations comprise multicast, in-network aggregation, or gradient compression.
Clause 6: The method of any of clauses 1-5, wherein the workload comprises one or more of a one-to-one transformation, one-to-many transformation, or many-to-one transformation.
Clause 7: The method of any of clauses 1-6, wherein the primitive is received by the network device via an application programming interface (API) operable to receive a message indicative of the primitive and instructions to the network device for executing the primitive.
Clause 8: The method of any of clauses 1-7, wherein the API is configured to receive a control path and a data path for a primitive.
Clause 9: The method of any of clauses 1-8, wherein the control path includes a transformation identifier, a match clause, and transformation information.
Clause 10: The method of any of clauses 1-9, wherein the match clause includes one of an offset from which to obtain values of the transformation identifier, or a Boolean valued filter expression.
Clause 11: A network device configured to process data packets in a computing network processing artificial intelligence (AI) workloads in a software defined network (SDN), the network device configured to execute primitives for packets associated with the AI workloads prior to forwarding the packets to hosts in the SDN, the network device configured to perform operations comprising:
Clause 12: The network device of clause 11, wherein the AI workloads comprise a collective.
Clause 13: The network device of any of clauses 11 and 12, wherein the operations comprise multicast, in-network aggregation, or gradient compression.
Clause 14: The network device of any of clauses 11-13, wherein the AI workload comprises one or more of a one-to-one transformation, one-to-many transformation, or many-to-one transformation.
Clause 15: The network device of any of clauses 11-14, wherein the primitive is received by the network device via an application programming interface (API) operable to receive a message indicative of the primitive and instructions to the network device for executing the primitive
Clause 16: The network device of any of clauses 11-15, wherein the API is configured to receive a control path and a data path for a primitive.
Clause 17: The network device of any of clauses 11-16, wherein the control path includes a transformation identifier, a match clause, and transformation information.
Clause 18: The network device of any of clauses 11-17, wherein the match clause includes one of an offset from which to obtain values of the transformation identifier, or a Boolean valued filter expression.
Clause 19: A system for processing data packets in a computing network processing artificial intelligence (AI) workloads in a network device operating in a software defined network (SDN), the network device configured to execute primitives for packets associated with the AI workloads prior to forwarding the packets to hosts in the SDN, the network device configured to perform operations comprising:
Clause 20: The system of clause 19, wherein the primitive is received by the network device via an application programming interface (API) operable to receive a message indicative of the primitive and instructions to the network device for executing the primitive.
