| Internet-Draft | FARE using BGP | July 2024 |
| Xu | Expires 2 January 2025 | [Page] |
Large language models (LLMs) like ChatGPT have become increasingly popular in recent years due to their impressive performance in various natural language processing tasks. These models are built by training deep neural networks on massive amounts of text data, often consisting of billions or even trillions of parameters. However, the training process for these models can be extremely resource-intensive, requiring the deployment of thousands or even tens of thousands of GPUs in a single AI training cluster. Therefore, three-stage or even five-stage CLOS networks are commonly adopted for AI networks. The non-blocking nature of the network become increasingly critical for large-scale AI models. Therefore, adaptive routing is necessary to dynamically load balance traffic to the same destination over multiple ECMP paths, based on network capacity and even congestion information along those paths.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119].¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 2 January 2025.¶
Copyright (c) 2024 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Large language models (LLMs) like ChatGPT have become increasingly popular in recent years due to their impressive performance in various natural language processing tasks. These models are built by training deep neural networks on massive amounts of text data, often consisting of billions or even trillions of parameters. However, the training process for these models can be extremely resource-intensive, requiring the deployment of thousands or even tens of thousands of GPUs in a single AI training cluster. Therefore, three-stage or even five-stage CLOS networks are commonly adopted for AI networks. Furthermore, In rail-optimized CLOS topologies with standard GPU servers (HB domain of eight GPUs), the Nth GPUs of each server in a group of servers are connected to the Nth leaf switch, which provides higher bandwidth and non-blocking connectivity between the GPUs in the same rail. In rail-optimized topology, most traffic between GPU servers would traverse the intra-rail networks rather than the inter-rail networks.¶
The non-blocking nature of the network, especially the network for intra-rail communication, become increasingly critical for large-scale AI models. AI workloads tend to be extremely bandwidth-hungry and they usually generate a few elephant flows simultaneAously. If the traditional hash-based ECMP load-balancing was used without any optimization, it's highly possible to cause serious congestion and high latency in the network once multiple elephant flows are routed to the same link. Since the job completion time depends on worst-case performance, serious congestion will result in model training time longer than expected. Therefore, adaptive routing is necessary to dynamically load balance traffic to the same destination over multiple ECMP paths, based on network capacity and even congestion information along those paths. In other words, adaptive routing is a capacity-aware and even congestion-aware path selection algorithm.¶
Furthermore, to reduce the congestion risk to the maximum extent, the routing should be more granular if possible. Flow-granular adaptive routing still has a certain statistical possibility of congestion. Therefore, packet-granular adaptive routing is more desirable although packet spray would cause out-of-order delivery issue. A flexible reordering mechanism must be put in place（e.g., egress ToRs or the receiving servers). Recent optimizations for RoCE and newly invented transport protocols as alternatives to RoCE no longer require handling out-of-order delivery at the network layer. Instead, the message processing layer is used to address it.¶
To enable adaptive routing, no matter whether flow-granular or packet-granular adaptive routing, it is necessary to propagate network topology information, including link capacity and/or even available link capacity (i.e., link capacity minus link load) across the CLOS network. Therefore, it seems straightforward to use link-state protocols such as OSPF or ISIS as the underlay routing protocol in the CLOS network, instead of BGP, for propagating link capacity information and/or even available link capacity information. How to leverage OSPF or ISIS to achieve adaptive routing has been described in [I-D.xu-lsr-fare]. However, some data center network operators have been used to the use of BGP as the underlay routing protocol of data center networks [RFC7938]. Therefore, there is a need to leverage BGP to achieve adaptive routing as well.¶
[I-D.ietf-idr-link-bandwidth] has specified a way to perform weighted ECMP based on link bandwidths conveyed in the non-transitive link bandwith extended community. However, it is impractical to enable adaptive routing by directly using the non-transitive link bandwidth extended community due to the following constraints as mentioned in [I-D.ietf-idr-link-bandwidth].¶
"No more than one link bandwidth extended community SHALL be attached to a route. Additionally, if a route is received with link bandwidth extended community and the BGP speaker sets itself as next-hop while announcing that route to other peers, the link bandwidth extended community should be removed. The extended community is optional non-transitive."¶
Hence, this document defines a new extended community referred to as Path Bandwidth Extended Community and describes how to use this newly defined path bandwidth extended community to achieve adaptive routing.¶
Note that while adaptive routing especially at the packet-granular level can help reduce congestion between switches in the network, thereby achieving a non-blocking fabric, it does not address the incast congestion issue which is commonly experienced in last-hop switches that are connected to the receivers in many-to-one communication patterns. Therefore, a congestion control mechanism is always necessary between the sending and receiving servers to mitigate such congestion.¶
The Path Bandwidth Extended Community is used to indicate the minimum bandwith of the path towards the destination. It is an new IPv4 Address Specific Extended Community that can be transitive or non-transitive.¶
The value of the high-order octet of this extended type is either 0x01 or 0x41. The low-order octet of this extended type is TBD.¶
The Value field consists of two sub-fields:¶
Global Administrator sub-field: This sub-field contains the router ID of the advertising router that appends the path bandwidth extended community or updates the path bandwidth value of the existing path bandwidth extended community.¶
Local Administrator sub-field: This sub-field contains the path bandwidth value in IEEE floating point format with units of Gigabytes per second (GB/s).¶
+----+ +----+ +----+ +----+
| S1 | | S2 | | S3 | | S4 | (Spine)
+----+ +----+ +----+ +----+
+----+ +----+ +----+ +----+ +----+ +----+ +----+ +----+
| L1 | | L2 | | L3 | | L4 | | L5 | | L6 | | L7 | | L8 | (Leaf)
+----+ +----+ +----+ +----+ +----+ +----+ +----+ +----+
Figure 1¶
(Note that the diagram above does not include the connections between nodes. However, it can be assumed that leaf nodes are connected to every spine node in their CLOS topology.)¶
In a three-stage CLOS network as shown in Figure 1, also known as a leaf-spine network, each leaf node would establish eBGP sessions with all spine nodes.¶
All nodes are enabled for adaptive routing.¶
When a leaf node, such as L1, advertises the route to a specific IP prefix that it originates, it will attach a transitive path bandwidth extended community filled with a maximum bandwidth value.¶
Upon receiving the above advertisement, a spine node, such as S1, SHOULD determine the minimum value between the bandwidth of the link towards the advertising node (e.g., L1) and the value of the path bandwidth extended community carried in the received route, and then update the path bandwidth extended community with the above minimum value before readvertising that route to remote eBGP peers. Once S1 receives multiple equal-cost routes for a given prefix from multiple leaf nodes (e.g., L1 and L2 in the server multi-homing scenario), for each route, it SHOULD determine the minimum value between the bandwidth of the link towards the advertising node and the value of the path bandwidth extended community carried in the received route, and then use that minimum bandwidth value as a weight value for that route when performing weighted ECMP. When readvertising the route for that prefix to remote eBGP peers further, the path bandwidth extended community would be updated with the sum of the minimum bandwidth value of each route.¶
When a leaf node, such as L8, receives multiple equal-cost routes for that prefix from spine nodes (e.g., S1, S2, S3 and S4), for each route, it will determine the minimum value between the bandwidth of the link towards the advertising node and the value of the path bandwidth extended community carried in the received route, and then use that minimum bandwidth value as a weight value for that route when performing weighted ECMP.¶
Note that the weighted ECMP according to path bandwidth SHOULD NOT be performed unless all equal-cost routes for a given prefix carry the path bandwidth extended community.¶
=========================================
# +----+ +----+ +----+ +----+ #
# | L1 | | L2 | | L3 | | L4 | (Leaf) #
# +----+ +----+ +----+ +----+ #
# PoD-1 #
# +----+ +----+ +----+ +----+ #
# | S1 | | S2 | | S3 | | S4 | (Spine) #
# +----+ +----+ +----+ +----+ #
=========================================
=============================== ===============================
# +----+ +----+ +----+ +----+ # # +----+ +----+ +----+ +----+ #
# |SS1 | |SS2 | |SS3 | |SS4 | # # |SS1 | |SS2 | |SS3 | |SS4 | #
# +----+ +----+ +----+ +----+ # # +----+ +----+ +----+ +----+ #
# (Super-Spine@Plane-1) # # (Super-Spine@Plane-4) #
#============================== ... ===============================
=========================================
# +----+ +----+ +----+ +----+ #
# | S1 | | S2 | | S3 | | S4 | (Spine) #
# +----+ +----+ +----+ +----+ #
# PoD-8 #
# +----+ +----+ +----+ +----+ #
# | L1 | | L2 | | L3 | | L4 | (Leaf) #
# +----+ +----+ +----+ +----+ #
=========================================
Figure 2¶
(Note that the diagram above does not include the connections between nodes. However, it can be assumed that the leaf nodes in a given PoD are connected to every spine node in that PoD. Similarly, each spine node (e.g., S1) is connected to all super-spine nodes in the corresponding PoD-interconnect plane (e.g., Plane-1).)¶
For a five-stage CLOS network as illustrated in Figure 2, each leaf node would establish eBGP sessions with all spine nodes of the same PoD while each spine node would establish eBGP sessions with all super-spine nodes in the corresponding PoD-interconnect plane.¶
In rail-optimized topology, Intra-rail communication with high bandwidth requirements would be restricted to a single PoD. Inter-rail communication with relatively lower bandwidth requirements need to travel across PoDs through PoD-interconnect planes. Therefore, enabling adaptive routing only in PoD networks is sufficient. It's optional to perform adaptive routing for cross-PoD traffic.¶
When a leaf node, such as L1 in PoD-1, advertises the route for a specific IP prefix that it originates, it will attach a transitive path bandwidth extended community filled with a maximum bandwidth value.¶
Upon receiving the above route advertisement, a spine node, such as S1 in PoD-1, will determine the minimum value between the bandwidth of the link towards the advertising node (e.g., L1 in PoD-1) and the value of the path bandwidth extended community carried in the route, and then update the path bandwidth extended community with the above minimum value before readvertising that route to remote eBGP peers. Once S1 in PoD-1 receives multiple equal-cost routes for a given prefix from multiple leaf nodes (e.g., L1 and L2 in PoD-1 in the server multi-homing scenario), for each route, it will determine the minimum value between the bandwidth of the link towards the advertising node and the bandwidth value of the path bandwidth extended community carried in the route, and then use that minimum bandwidth value as a weight value for that route when performing weighted ECMP. When readvertising the route for that prefix to remote eBGP peers, the path bandwidth extended community would be updated with the sum of the minimum bandwidth value of each route.¶
When a given super-spine node, such as SS1 in Plane-1, receives the route for that prefix from S1 in PoD-1, it will not update the transtive path bandwidth extended community when readvertising that route. It COULD optionally attach another path bandwidth extended community which is non-transitive to indicate the bandwith of the link towards the advertising router.¶
When a given spine node in another PoD, such as S1 in PoD-8, receives multiple equal-cost routes for a given prefix from super-spine nodes in Plane-1 (e.g., SS1, SS2, SS3 and SS4 in Plane-1), it will not update the value of the transitive path bandwidth extended community when readvertising that route towards remote peers (Note that the transitive path bandwidth extended community of those multiple equal-cost routes carry the same value that was set by S1 in PoD-1). Meanwhile, once each route contains a non-transitive path bandwidth extended community, for each route, it will determine the minimum value between the bandwidth of the link towards the advertising node and the bandwidth value of the non-transitive path bandwidth extended community carried in the route, and then use that minimum bandwidth value as a weight value for that route when performing weighted ECMP.¶
When a leaf node, such as L8 in PoD-8, receives multiple equal-cost routes for that prefix from multiple spine nodes (e.g., S1, S2, S3 and S4 in PoD-8), for each route, it will determine the minimum value between the bandwidth of the link towards the advertising node and the value of the path bandwidth extended community carried in the route, and then use that minimum bandwidth value as a weight value for that route when performing weighted ECMP.¶
Note that the weighted ECMP according to path bandwidth SHOULD NOT be performed unless all equal-cost routes for a given prefix carry the path bandwidth extended community.¶
TBD.¶
TBD.¶