Network virtualization over infiniband

Abstract

Mechanisms are provided to allow servers connected over an InfiniBand fabric to communicate using multiple private virtual interconnects (PVIs). In particular embodiments, the PVIs appear as virtual Ethernet networks to users on individual servers and virtual machines running on the individual servers. Each PVI is represented on the server by a virtual network interface card (VNIC) and each PVI is mapped to its own InfiniBand multicast group. Data can be transmitted on PVIs as Ethernet packets fully encapsulated, including the layer 2 header, within InfiniBand messages. Broadcast and multicast frames are propagated using InfiniBand.

Description

TECHNICAL FIELD

The present disclosure relates to network virtualization over InfiniBand.

DESCRIPTION OF RELATED ART

InfiniBand provides a robust, scalable, and fail-safe architecture for connecting nodes such as servers, appliances, and disk arrays. InfiniBand is often used in high performance server clusters and datacenters. In one particular application, InfiniBand is used to connect servers to an input/output (I/O) director that provides efficient virtualized, shared, and fault tolerant I/O resources such as host bus adapters (HBAs) and network interface cards (NICs) to the servers.

However, mechanisms for isolating or separating communications on an InfiniBand fabric are limited. Furthermore, other mechanisms such as Internet Protocol (IP) over InfiniBand (IB) do not easily allow for efficient virtualization. Consequently, techniques and mechanisms are provided to enhance communications over InfiniBand and allow for network virtualization over InfiniBand.

BRIEF DESCRIPTION OF THE DRAWINGS

The disclosure may best be understood by reference to the following description taken in conjunction with the accompanying drawings, which illustrate particular example embodiments.

FIG. 1 illustrates one example of a system with servers connected to an I/O director.

FIG. 2 illustrates one example of a system having multiple servers and multiple private virtual interconnects (PVIs) over InfiniBand.

FIG. 3 illustrates one example of a technique for creating a PVI.

FIG. 4 illustrates one example of a forwarding table.

FIG. 5 illustrates one example of a technique for sending data.

FIG. 6 provides one example of a system that can be used to implement one or more mechanisms.

DESCRIPTION OF PARTICULAR EMBODIMENTS

Reference will now be made in detail to some specific examples of the invention including the best modes contemplated by the inventors for carrying out the invention. Examples of these specific embodiments are illustrated in the accompanying drawings. While the invention is described in conjunction with these specific embodiments, it will be understood that it is not intended to limit the invention to the described embodiments. On the contrary, it is intended to cover alternatives, modifications, and equivalents as may be included within the spirit and scope of the invention as defined by the appended claims.

For example, the techniques and mechanisms of the present invention will be described in the context of InfiniBand and an input/output (I/O) director. However, it should be noted that the techniques and mechanisms of the present invention apply to InfiniBand variations and other types of networks as well as architectures that do not include an I/O director. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. Particular example embodiments of the present invention may be implemented without some or all of these specific details. In other instances, well known process operations have not been described in detail in order not to unnecessarily obscure the present invention.

Various techniques and mechanisms of the present invention will sometimes be described in singular form for clarity. However, it should be noted that some embodiments include multiple iterations of a technique or multiple instantiations of a mechanism unless noted otherwise. For example, a system uses a processor in a variety of contexts. However, it will be appreciated that a system can use multiple processors while remaining within the scope of the present invention unless otherwise noted. Furthermore, the techniques and mechanisms of the present invention will sometimes describe a connection between two entities. It should be noted that a connection between two entities does not necessarily mean a direct, unimpeded connection, as a variety of other entities may reside between the two entities. For example, a processor may be connected to memory, but it will be appreciated that a variety of bridges and controllers may reside between the processor and memory. Consequently, a connection does not necessarily mean a direct, unimpeded connection unless otherwise noted.

OVERVIEW

Mechanisms are provided to allow servers connected over an InfiniBand fabric to communicate using multiple private virtual interconnects (PVIs). In particular embodiments, the PVIs appear as virtual Ethernet networks to users on individual servers and virtual machines running on the individual servers. Each PVI is represented on the server by a virtual network interface card (VNIC) and each PVI is mapped to its own InfiniBand multicast group. Data can be transmitted on PVIs as Ethernet packets fully encapsulated including the layer 2 header within InfiniBand messages. Broadcast and multicast frames are propagated using InfiniBand.

Example Embodiments

InfiniBand is a switched fabric that provides high bandwidth, low latency, quality of service, and failover capabilities. InfiniBand provides point-to-point bidirectional serial links to connect servers, disk arrays, appliances, etc. InfiniBand offers unicast, multicast, and broadcast support and is often used in cloud computing clusters and data centers.

In particular embodiments, the servers are connected over an InfiniBand fabric to an I/O director. The I/O director provides shared and virtualized I/O resources to the servers. The common approach for providing I/O connectivity to servers and other hosts is to provide I/O controllers within the servers themselves. I/O controllers include Ethernet network interface cards (NICs), Fibre Channel, iSCSI and SAS host bus adapters (HBAs), etc. The I/O controllers are then connected to external devices using cables. External devices include switches, storage devices, display devices, and others. Cabling quickly becomes hard to manage in data centers with a significant number of servers, networks, and storage devices.

In some implementations, I/O controllers are offloaded onto an external shared system referred to herein as an I/O director. The I/O director includes actual I/O resources connected to external devices such as switches and storage. The hosts are connected to the I/O director over InfiniBand, but the number of cables required to provide redundancy and fault tolerance is much lower than the number of cables required when each host has its own I/O resources. In many cases, deploying an I/O director reduces the number of I/O cables per server from half a dozen or a dozen to one or two cables. A VNIC driver is provided for communication with the VNIC I/O modules at the I/O director and for providing network device services on the server which correspond to those provided by local physical NICs. The end result is that servers have connectivity to any number of different data and storage networks using virtual I/O devices.

Although servers can efficiently communicate with external entities on external networks using virtualized I/O resources, communications with other servers on the same InfiniBand fabric are not necessarily efficient. Communications with other servers on the same InfiniBand fabric are still required to pass through the I/O module at the I/O director. Having local communications pass through the I/O module at the I/O director is inefficient and introduces a significant bandwidth, latency, and throughput limitations. Furthermore, if virtual networks are desired, one port at the I/O module is required for each separate virtual network. This can be problematic in systems that require thousands of virtual networks.

Consequently, the techniques of the present invention provide mechanisms for implementing virtual networks in an InfiniBand fabric. According to various embodiments, servers are connected over an InfiniBand fabric using virtual NICs (VNICs) that encapsulate Ethernet packets including layer 2 headers in InfiniBand messages. Servers and virtual machines can communicate as though the servers and virtual machines are connected using an Ethernet architecture. Different VNICs are provided for each virtual network. According to various embodiments, each virtual network is referred to herein as a private virtual interconnect (PVI). Each PVI provides logically isolated communications. A server may be a member of any number of PVIs.

According to various embodiments, an administrator uses a management system to assign PVIs to servers. Each PVI is represented on the server by a VNIC which is used to access the PVI. Each PVI is mapped to its own InfiniBand multicast group which serves as its broadcast domain. PVI unicast frames are encapsulated in their entirety within InfiniBand reliable connected (RC) and unreliable datagram (UD) protocol messages. By contrast, a mechanism such as IP over InfiniBand does not include layer 2 headers in encapsulation.

Broadcast and multicast frames are propagated using InfiniBand multicast operations. In particular embodiments, mechanisms are provided for learning mappings between layer 2 addresses used within the PVI and their corresponding InfiniBand end points. Failover in case of link or switch failure is supported.

According to various embodiments, a very large number of isolated virtual networks can be created and scaled in a manner that allows high performance server-to-server communication. The mechanism is scalable, easy to manage, and provides significant benefits for a variety of applications. In particular embodiments, all intelligence can be maintained within VNIC drivers at individual servers. No centralized controller is required. Discovery can be performed autonomously with existing InfiniBand messages. Users on servers and virtual machines have access to their own networks that appear to them as Ethernet networks.

FIG. 1 illustrates one example of a system that includes multiple servers connected using an InfiniBand fabric to an I/O director. In particular embodiments, multiple servers 101, 103, 105, 107, and 109 are linked through an interconnect 131 such as an InfiniBand fabric. According to various embodiments, the servers 101, 103, 105, 107, and 109 communicate using Ethernet packets encapsulated in InfiniBand messages. VNICs 111, 115, 119, 123, and 127 are provided for servers 101, 103, 105, 107, and 109 respectively. According to various embodiments, VNICs 111, 115, 119, 123, and 127 are virtual network interface cards that appear to users at individual servers to be actual network interface cards.

To communicate with entities on an external network 161, servers 101, 103, 105, 107, and 109 use VNICs 111, 115, 119, 123, and 127 respectively to communicate with an I/O director 151 over the InfiniBand fabric. According to various embodiments, the I/O director 151 includes I/O ports 141. I/O ports 141 include VNICs that provide the servers 101, 103, 105, 107, and 109 with virtualized I/O resources. According to various embodiments, the I/O director includes a target channel adapter (TCA) for actual communications on the InfiniBand fabric. A TCA can be a discrete device, or its functionality can be integrated into another device of the I/O module. A TCA may recognize and terminate various transport protocols (iWARP, RC, etc.)

According to various embodiments, the TCA removes the link and transport protocol headers from the packet when a server transmits a data packet to the I/O ports 141. The TCA then forwards the packet with an internal header to a network processor in the I/O director 151.

According to various embodiments, the network processor may include VNIC to VNIC switching logic. The VNIC-to-VNIC switching logic performs packet forwarding between VNICs terminating on the same Ethernet port. The VNIC-to-VNIC switching logic 227 maintains a table of corresponding VNICs and MAC addresses and performs packet forwarding based on MAC addresses. For example, if VNIC_—1 is linked to address MAC_—1, and a data packet having MAC_—1 as its destination address is received on VNIC_—2 which terminates on the same Ethernet port as VNIC 1, then the VNIC-to-VNIC switching logic forwards this packet to VNIC_—1. This functionality allows use of an I/O director with external switches that do not forward packets to the same link that they came from, so that the switching is performed, in this case, within I/O modules themselves.

According to various embodiments, the VNIC I/O module also has learning logic, which is used to establish a mapping of VNICs created by virtualization software (on the servers) to VNICs of the I/O director. When a server is virtualized and one or more virtual machines are created on the server, each virtual machine can be associated with one or more VNICs, which are implemented by the server virtualization software. These VNICs are also referred to as virtual machine VNICs or simply VM VNICs. According to various embodiments, each VM VNIC has a MAC address, which is assigned by the virtualization software. One or more VM VNICs may be bridged to a single VNIC of the I/O director using a software virtual switch, which is implemented by the virtualization software. In particular embodiments, the traffic of multiple VM VNICs may appear on the same VNIC of the I/O director, and this traffic may include packets with different source MAC addresses for the different VM VNICs. According to various embodiments, the VNIC I/O module 203 establishes a mapping between a VM VNIC MAC address and a corresponding VNIC of the I/O director. This mapping enables directing incoming traffic to the correct VNIC of the I/O director. For example, if a packet with destination MAC address MAC_—1 arrives at the I/O module Ethernet port, and MAC_—1 is the address of VM VNIC_—1, then the I/O module needs to know which VNIC of the I/O director should receive this packet. In certain embodiments, a lookup is performed in a mapping table to establish this I/O director VNIC to VM VNIC correspondence.

By using VNICs 111, 115, 119, 123, and 127 to communicate with an I/O director 151, communication with external network 161 can be performed efficiently using shared and virtualized I/O resources. However, even communications between servers 101, 103, 105, 107, and 109 that are not destined for any external network 161 have to go through the I/O director 151. Requiring all inter-server communications to go through the I/O director 151 is inefficient and introduces an artificial bottleneck into a system. Furthermore, the number of virtual networks that can be created in an InfiniBand fabric is limited by the number of ports in an I/O module of the I/O director 151. For example, creating 1500 virtual networks in an InfiniBand fabric would require 1500 ports.

FIG. 2 illustrates one example of system that includes multiple servers connected over multiple virtual networks. An InfiniBand fabric includes multiple servers 201, 203, 205, 207, and 209. According to various embodiments, server 201 is assigned VNIC1211 and VNIC2213. Server 203 is assigned to VNIC1215 and VNIC2217. Server 205 is assigned VNIC1219, VNIC2221 and VNIC3223. In particular embodiments, server 207 is a load balancer or other appliance assigned VNIC2225 and VNIC3227. Server 209 is assigned VNIC2229 and VNIC3231.

According to various embodiments, servers 201, 203, and 205 assigned VNIC1211, 215, and 219, respectively, are members of private virtual interconnect (PVI) 241. Servers 201, 203, 205, 207, and 209 assigned VNIC2213, 217, 221, 225, and 229, respectively, are members of PVI 243. Servers 205, 207, and 209 assigned VNIC3223, 227, and 231, respectively, are members of PVI 245. According to various embodiments, communications on PVI 241, 243, and 245 are transmitted as Ethernet packets including layer 2 headers encapsulated in InfiniBand reliable connected (RC) and unreliable datagram (UD) protocol messages. According to various embodiments, a PVI can be created when an administrator directs a server to create a new VNIC corresponding to a virtual network identifier such as a net_ID. According to various embodiments, the net_ID is translated to a multicast group identifier by performing minor bit modification. Based on multicast group identifier, a multicast group join operation is propagated to the subnet manager.

If the server is the first member of the multicast group corresponding to a virtual network, the subnet manager creates a multicast group and adds the port to the multicast group using the multicast group ID corresponding to the net_ID and programs all switches on the path to add the new port. If the server is not the first member, the subnet manager adds the port to the multicast group and programs all switches on the path to add the new port. A driver then creates the VNIC on the server. It should be noted that InfiniBand elements, such as queue pairs necessary for communication, may also be created at this point, e.g. for UD communications or later on, e.g. for RC communications. A queue pair may include a send queue and a receive queue created in tandem and identified by a queue pair number.

FIG. 3 illustrates one example of a mechanism for creating one or more private virtual interconnects (PVIs) in an InfiniBand network connecting multiple servers and/or appliances such as load balancers and security systems. The servers may or may not be connected to an I/O director that provides shared and virtualized I/O resources to the servers. According to various embodiments, an instruction is received at 301 to include a server in a virtual network. At 303, a net_ID corresponding to the virtual network is identified. At 305, the net_ID is translated to a multicast group ID using minor bit modification. Based on multicast group identifier, a multicast group join operation is propagated to the subnet manager. If the server is the first member of the multicast group corresponding to a virtual network, the subnet manager creates at 309 a multicast group and adds the port to the multicast group using the multicast group ID corresponding to the net_ID and programs all switches on the path to add the new port. If the server is not the first member,

the subnet manager creates a multicast group and adds the port to the multicast group using the multicast group ID corresponding to the net_ID and programs all switches on the path to add the new port at 309. If the server is not the first member, the subnet manager adds the port to the multicast group and programs all switches on the path to add the new port at 311. According to various embodiments, a driver then creates the new VNIC on the server at 315.

FIG. 4 illustrates one example of a forwarding table used for transmitting data in an InfiniBand network that supports multiple virtual networks. A private virtual interconnect driver forwarding table 401 is provided on a per VNIC driver basis. The forwarding table 401 includes a destination address 411, a VLAN identifier 413, destination InfiniBand address information 415, and destination queue pair information 417. According to various embodiments, the destination InfiniBand address information 415 may be a destination InfiniBand address vector. In particular embodiments, the destination 411 and VLAN ID 413 pair are used to identify unique forwarding table entries. The destination InfiniBand address info 415 and destination queue pair 417 are used to forward data based on InfiniBand standard UD and RC mechanisms.

FIG. 5 illustrates one example of a technique for sending data. According to various embodiments, a VNIC driver receives data from a network stack at 501. The data may be Ethernet data that the VNIC driver encapsulates in an InfiniBand message at 503. It is determined whether the data corresponds to a broadcast packet, a multicast packet, or a unicast packet at 505. If the data corresponds to a broadcast packet, a multicast encapsulated packet is transmitted on the PVI queue pair at 507. If the data corresponds to a multicast packet, an InfiniBand multicast group is identified at 509. In some instances, the multicast packet can be treated as a broadcast packet and transmitted to everyone on the PVI multicast group. In other instances, an InfiniBand multicast group is created for each IP multicast group is used for multicast operations. The multicast group packet can then be transmitted using the IB multicast group at 511.

If the data corresponds to a unicast packet, the destination address is accessed in a forwarding table specific to that VNIC driver at 513. The destination address and a VLAN ID are used to identify a unique entry in the forwarding table 515. Conventional InfiniBand forwarding mechanisms are then used to transmit UD and RC packets.

When a destination server receives an InfiniBand message, InfiniBand message encapsulation is removed to extract Ethernet data. Information from the InfiniBand message can be used to populate a forwarding table at the destination server. Information may include destination queue pair and destination address.

According to various embodiments, the various mechanisms can be implemented in hardware, firmware, and/or software. FIG. 6 provides one example of a system that can be used to implement one or more mechanisms. For example, the system shown in FIG. 6 may be used to implement a server or an I/O director.

According to particular example embodiments, a system 600 suitable for implementing particular embodiments of the present invention includes a processor 601, a memory 603, an interface 611, and a bus 615 (e.g., a PCI bus). When acting under the control of appropriate software or firmware, the processor 601 is responsible for such tasks such as data modification. Various specially configured devices can also be used in place of a processor 601 or in addition to processor 601. The complete implementation can also be done in custom hardware. The interface 611 is typically configured to send and receive data packets or data segments over a network. Particular examples of interfaces the device supports include host bus adapter (HBA) interfaces, Ethernet interfaces, frame relay interfaces, cable interfaces, DSL interfaces, token ring interfaces, and the like.

In addition, various very high-speed interfaces may be provided such as fast Ethernet interfaces, 1/10/40/100 Gigabit Ethernet interfaces, ATM interfaces, HSSI interfaces, POS interfaces, FDDI interfaces, Host Channel Adapter, and the like. Generally, these interfaces may include ports appropriate for communication with the appropriate media. In some cases, they may also include an independent processor and, in some instances, volatile RAM. The independent processors may control communications-intensive tasks.

According to particular example embodiments, the system 600 uses memory 603 to store data, algorithms, and program instructions. The program instructions may control the operation of an operating system and/or one or more applications, for example. The memory or memories may also be configured to store received data and process received data.

Because such information and program instructions may be employed to implement the systems/methods described herein, the present invention relates to tangible, machine readable media that include program instructions, state information, etc. for performing various operations described herein. Examples of machine-readable media include, but are not limited to, magnetic media such as hard disks, floppy disks, and magnetic tape; optical media such as CD-ROM disks and DVDs; magneto-optical media such as optical disks; and hardware devices that are specially configured to store and perform program instructions, such as read-only memory devices (ROM) and random access memory (RAM). Examples of program instructions include both machine code, such as produced by a compiler, and files containing higher level code that may be executed by the computer using an interpreter.

Although the foregoing invention has been described in some detail for purposes of clarity of understanding, it will be apparent that certain changes and modifications may be practiced within the scope of the appended claims. Therefore, the present embodiments are to be considered as illustrative and not restrictive and the invention is not to be limited to the details given herein, but may be modified within the scope and equivalents of the appended claims.

Claims

1. A method for creating a virtual network, the method comprising: converting a virtual network identifier to an InfiniBand multicast group identifier at a first server;sending an InfiniBand multicast message over an InfiniBand fabric, the InfiniBand fabric including the first server, a second server, and a third server, wherein communications between the first server, the second server, and the third server comprise Ethernet packets encapsulated for transmission over the InfiniBand fabric, wherein a network entity receives the InfiniBand multicast message and determines whether the first server is a first member of a multicast group corresponding to a virtual network and adds a port to a multicast group;creating a virtual network interface card (VNIC) corresponding to the virtual network identifier for each of the first server, second server, and third server; andcreating a Private Virtual Interconnect (PVI) between two or more of the first server, second server, or third server using the VNIC, the PVI comprising a virtual Ethernet network corresponding to the virtual network identifier, wherein the two or more of the first, second, and third servers of the virtual Ethernet network communicate via the PVI using Ethernet packets encapsulated within InfiniBand messages.

2. The method of claim 1, wherein the InfiniBand multicast message is associated with an InfiniBand multicast join operation.

3. The method of claim 1, wherein the network entity is a subnet manager.

4. The method of claim 3, wherein if the first server is the first member of the multicast group, the subnet manager creates the multicast group.

5. The method of claim 1, wherein the multicast group corresponds to the virtual network.

6. The method of claim 1, wherein converting the virtual network identifier to the InfiniBand multicast group comprises performing bit modification.

7. The method of claim 1, wherein the VNIC is created by a driver on the first server.

8. The method of claim 1, wherein the second server is a network appliance.

9. A system comprising: a processor; anda memory coupled with and readable by the processor and storing therein a set of instructions which, when executed by the processor, causes the processor to create a virtual network by: converting a virtual network identifier to an InfiniBand multicast group identifier at a first server;sending an InfiniBand multicast message over an InfiniBand fabric, the InfiniBand fabric including the first server, a second server, and a third server, wherein communications between the first server, the second server, and the third server comprise Ethernet packets encapsulated for transmission over the InfiniBand fabric, wherein a network entity receives the InfiniBand multicast message and determines whether the first server is a first member of a multicast group corresponding to a virtual network and adds a port to a multicast group;creating a virtual network interface card (VNIC) corresponding to the virtual network identifier for each of the first server, second server, and third server; andcreating a Private Virtual Interconnect (PVI) between two or more of the first server, second server, or third server using the VNIC, the PVI comprising a virtual Ethernet network corresponding to the virtual network identifier, wherein the two or more of the first, second, and third servers of the virtual Ethernet network communicate via the PVI using Ethernet packets encapsulated within InfiniBand messages.

10. The system of claim 9, wherein the InfiniBand multicast message is associated with an InfiniBand multicast join operation.

11. The system of claim 9, wherein the network entity is a subnet manager.

12. The system of claim 11, wherein if the first server is the first member of the multicast group, the subnet manager creates the multicast group.

13. The system of claim 9, wherein the multicast group corresponds to a virtual network.

14. The system of claim 9, wherein converting the virtual network identifier to the InfiniBand multicast group comprises performing bit modification.

15. The system of claim 9, wherein the VNIC is created by a driver on the first server.

16. The system of claim 9, wherein the second server is a network appliance.

17. A non-transitory computer readable medium comprising a set of instructions stored therein which, when executed by a processor, causes the processor to create a virtual network by: converting a virtual network identifier to an InfiniBand multicast group identifier at a first server;sending an InfiniBand multicast message over an InfiniBand fabric, the InfiniBand fabric including the first server, a second server, and a third server, wherein communications between the first server, the second server, and the third server comprise Ethernet packets encapsulated for transmission over the InfiniBand fabric, wherein a network entity receives the InfiniBand multicast message and determines whether the first server is a first member of a multicast group corresponding to a virtual network and adds a port to a multicast group;creating a virtual network interface card (VNIC) corresponding to the virtual network identifier for each of the first server, second server, and third server; andcreating a Private Virtual Interconnect (PVI) between two or more of the first server, second server, or third server using the VNIC, the PVI comprising a virtual Ethernet network corresponding to the virtual network identifier, wherein the two or more of the first, second, and third servers of the virtual Ethernet network communicate via the PVI using Ethernet packets encapsulated within InfiniBand messages.

18. The non-transitory computer readable medium of claim 17, wherein the InfiniBand multicast message is associated with an InfiniBand multicast join operation.