# US 2019/0173810 A1 · Packet scheduling in a switch for reducing cache-miss rate at a destination network node

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2019/0173810 A1
- Document ID: US20190173810A1
- Status: Published application
- Kind: A1
- Application number: 15/832,806
- Filed: 2017-12-06
- Published: 2019-06-06
- Earliest claimed: 2017-12-06
- Assignees: Mellanox Technologies Tlv Ltd.
- Applicants: Mellanox Technologies TLV Ltd.
- Inventors: Alex Shpiner; Tal Anker; Matty Kadosh
- Web page: https://patentgrep.com/patent/US20190173810A1

## CPC

- H04L 49/9057: Arrangements for supporting packet reassembly or resequencing
- G06F 15/17331: {Distributed shared memory [DSM], e.g. remote direct memory access [RDMA]}
- H04L 47/6225: {Fixed service order, e.g. Round Robin}
- H04L 47/624: {Altering the ordering of packets in an individual queue}
- H04L 67/1097: for distributed storage of data in networks, e.g. transport arrangements for network file system [NFS], storage area networks [SAN] or network attached storage [NAS]
- H04L 67/568: Storing data temporarily at an intermediate stage, e.g. caching
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A network switch includes switch circuitry and multiple ports. The ports are configured to communicate with a communication network. The switch circuitry is configured to receive via the ports multiple packets, which are destined to a destination network node and which specify attributes used by the destination network node as cache keys for on-demand fetching of context items into a cache memory of the destination network node, to control a rate of fetching the context items into the cache memory at the destination network node, by ordering the received packets in a sequence, based on the attributes of the respective packets, using an ordering criterion that aims to place packets that access a common context item in proximity to one another in the sequence, and to forward the received packets to the destination network node, via the ports, in accordance with the ordered sequence.

## Claims

### Claim 1

A network switch, comprising:
multiple ports configured to communicate with a communication network; and
switch circuitry configured to:
receive via the ports multiple packets, which are destined to a destination network node and which specify attributes used by the destination network node as cache keys for on-demand fetching of context items into a cache memory of the destination network node;
control a rate of fetching the context items into the cache memory at the destination network node, by ordering the received packets in a sequence, based on the attributes of the respective packets, using an ordering criterion that aims to place packets that access a common context item in proximity to one another in the sequence; and
forward the received packets to the destination network node, via the ports, in accordance with the ordered sequence.

### Claim 2 (depends on 1)

The network switch according to claim 1, further comprising multiple queues for storing packets awaiting transmission to the communication network, wherein the switch circuitry is configured to order the received packets by distributing the received packets among the multiple queues so that at least some of the received packets that are associated with a same cache key are distributed to a common queue.

### Claim 3 (depends on 2)

The network switch according to claim 2, wherein the switch circuitry is configured to scan the multiple queues and to read from each queue visited multiple packets, if available, before switching to read a subsequent queue among the multiple queues.

### Claim 4 (depends on 2)

The network switch according to claim 2, wherein the switch circuitry is configured to read the multiple queues in accordance with a Deficit Round Robin (DRR) schedule scheme that assigns to each of the multiple queues a quanta parameter of multiple packets.

### Claim 5 (depends on 2)

The network switch according to claim 2, wherein the received packets are associated with multiple respective priority levels, wherein the switch circuitry is configured to distribute the received packets to multiple respective queue-groups that each comprises two or more of the multiple queues, based on the respective priority levels, and to read multiple packets from a selected queue in one queue-group before switching to selecting a subsequent queue for reading in another queue-group.

### Claim 6 (depends on 2)

The network switch according to claim 2, wherein the switch circuitry is configured to receive additional packets that are destined to the destination network node, wherein the additional packets require no access to any cached context item at the destination network node, and to distribute the additional packets to a dedicated queue that is separate from the queues allocated for storing packets to be ordered.

### Claim 7 (depends on 6)

The network switch according to claim 6, wherein the switch circuitry is configured to read the queues by alternating reading one or more packets from a selected queue among the queues storing the packets to be ordered, and one or more packets from the dedicated queue.

### Claim 8 (depends on 1)

The network switch according to claim 1, wherein the context items in the cache memory comprise information for translating between destination virtual addresses carried in headers of the received packets and physical addresses of a memory of the destination network node in which a payload part of the received packets is to be stored, and wherein the attributes used by the destination network node as the cache keys comprise the virtual addresses.

### Claim 9 (depends on 1)

The network switch according to claim 1, wherein the received packets are organized in Remote Direct Memory Access (RDMA) messages for directly accessing data in the memory of the destination network node.

### Claim 10 (depends on 9)

The network switch according to claim 9, wherein the communication network comprises an InfiniBand fabric, and wherein the RDMA messages comprise messages communicated in accordance with a RDMA over Converged Ethernet (RoCE) protocol.

### Claim 11

A method for communication, comprising:
in a network switch that comprises multiple ports for communicating with a communication network, receiving via the ports multiple packets, which are destined to a destination network node and which specify attributes used by the destination network node as cache keys for on-demand fetching of context items into a cache memory of the destination network node;
controlling a rate of fetching context items into the cache memory at the destination network node, by ordering the received packets in a sequence, by the network switch, based on the attributes of the respective packets, using an ordering criterion that aims to place packets that access a common context item in proximity to one another in the sequence; and
forwarding the packets to the destination network node, via the ports, in accordance with the ordered sequence.

### Claim 12 (depends on 11)

The method according to claim 11, wherein the network switch further comprises multiple queues for storing packets awaiting transmission to the communication network, wherein ordering the received packets comprises distributing the received packets among the multiple queues so that at least some of the received packets that are associated with a same cache key are distributed to a common queue.

### Claim 13 (depends on 12)

The method according to claim 12, wherein distributing the received packets comprises scanning the multiple queues and reading from each queue visited multiple packets, if available, before switching to read a subsequent queue among the multiple queues.

### Claim 14 (depends on 12)

The method according to claim 12, wherein ordering the received packets comprises reading the multiple queues in accordance with a Deficit Round Robin (DRR) schedule scheme that assigns to each of the multiple queues a quanta parameter of multiple packets.

### Claim 15 (depends on 12)

The method according to claim 12, wherein the received packets are associated with multiple respective priority levels, wherein ordering the received packets comprises distributing the received packets to multiple respective queue-groups that each comprises two or more of the multiple queues, based on the respective priority levels, and reading multiple packets from a selected queue in one queue-group before switching to selecting a subsequent queue for reading in another queue-group.

### Claim 16 (depends on 12)

The method according to claim 12, and comprising receiving additional packets that are destined to the destination network node, wherein the additional packets require no access to any cached context item at the destination network node, and distributing the additional packets to a dedicated queue that is separate from the queues allocated for storing packets to be ordered.

### Claim 17 (depends on 16)

The method according to claim 16, wherein ordering the received packets comprises reading the queues by alternating reading one or more packets from a selected queue among the queues storing the packets to be ordered, and one or more packets from the dedicated queue.

### Claim 18 (depends on 11)

The method according to claim 11, wherein the context items in the cache memory comprise information for translating between destination virtual addresses carried in headers of the received packets and physical addresses of a memory of the destination network node in which a payload part of the received packets is to be stored, and wherein the attributes used by the destination network node as the cache keys comprise the virtual addresses.

### Claim 19 (depends on 11)

The network switch according to claim 11, wherein the received packets are organized in Remote Direct Memory Access (RDMA) messages for directly accessing data in the memory of the destination network node.

### Claim 20 (depends on 19)

The method according to claim 19, wherein the communication network comprises an InfiniBand fabric, and wherein the RDMA messages comprise messages communicated in accordance with a RDMA over Converged Ethernet (RoCE) protocol.
