Browse Prior Art Database

INTELLIGENT LOAD BALANCING OVER FABRICPATH IN DATA CENTER

IP.com Disclosure Number: IPCOM000240717D
Publication Date: 2015-Feb-20
Document File: 9 page(s) / 167K

Publishing Venue

The IP.com Prior Art Database

Related People

Amir Khan: AUTHOR

Abstract

The health/utilization of the uplinks in a data center is considered before they can be selected for load balancing. Hashing alone would not be able to select the best link and can lead to congestion. This solution minimizes the use of hashing to the first packet of a particular flow. No more hashing is required once the Flow-Link mapping is identified.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 23% of the total text.

Page 01 of 9

INTELLIGENT LOAD BALANCING OVER FABRICPATH IN DATA CENTER

AUTHORS:

Amir Khan

CISCO SYSTEMS, INC.

ABSTRACT

    The health/utilization of the uplinks in a data center is considered before they can be selected for load balancing. Hashing alone would not be able to select the best link and can lead to congestion. This solution minimizes the use of hashing to the first packet of a particular flow. No more hashing is required once the Flow-Link mapping is identified.

DETAILED DESCRIPTION

    In current data center deployments, load balancing between the Access layer and the Aggregation layer has become very critical, both for North-South traffic as well as for East-West traffic. There are various load balancing technologies available, such as Port- Channel, Emulated Switch, Hot Standby Router Protocol (HSRP) anycast etc. All these technologies are just a variation of Equal Cost Multi-Path (ECMP) at Layer 2.

    Figure 1 below shows a network including the Access layer and the Aggregation layer, between which is a L2 FabricPath network. There can more aggregation layer switches. North-South traffic originates behind Access layer switches and goes towards the core of data center. Tis traffic is load balanced using hashing by the Access layer switches.

Copyright 2015 Cisco Systems, Inc.

1


Page 02 of 9

Figure 1

    Port-Channel creates a logical channel with a bunch of physical ports mapped to it. Emulated Switch creates 2 ECMP paths for a logical switch shared between two real switches. HSRP anycast can achieve load balancing for up to 32 ECMP paths. All these technologies are dependent on one big constraint to do a proper load balancing and achieve the desired result of using all available ECMPs. That is, the hash value is calculated by the Access layer switch. The hash value defines which uplink the Aggregation layer selects among multiple ECMPs. This hash is generally calculated based on SrcIP/DestIP/SrcPort/DestPort of the outgoing packet (although there is a Command Line Interface to adjust these hash parameters and select others parameters like SrcMAC/DestMAC etc.). The Access layer switch takes these values from the

Copyright 2015 Cisco Systems, Inc.

2


Page 03 of 9

packet, generates a hash and uses the hash to select the uplink. This is a common way of selecting uplinks across all the above-mentioned technologies, but also in VxLAN.

    In VxLAN all the packets flowing between a pair of VTEPs have the same SrcIP/DestIP/DestPort. Only the Source port (which is derived out of a hash done on inner packet L2 frame source MAC and destination MAC) which is different between different flows and then again this combination of SrcIP/DestIP/DestPort/SrcPort is passed through a hash function to select the uplink.

    There is basic problem with the above approach of choosing the uplink. Specifically, what if there is a hash collision for multiple flows? This could result in a particular uplink being over-utilized and others remain under-utilized. The Aggregation lay...