[Idr] 答复: [RTGWG & IDR] Discussing the Open Scheduled fabric of AI Network

Aijun Wang <wangaijun@tsinghua.org.cn> Fri, 05 July 2024 03:00 UTC

From: Aijun Wang <wangaijun@tsinghua.org.cn>
To: '霍朋飞' <huopengfei=40bytedance.com@dmarc.ietf.org>, idr@ietf.org, rtgwg@ietf.org
References: <CA++4kv88HEKoA2Mvwh0nUyM-uyR97UhDh=w9jOGVfMjASdnw8w@mail.gmail.com>
In-Reply-To: <CA++4kv88HEKoA2Mvwh0nUyM-uyR97UhDh=w9jOGVfMjASdnw8w@mail.gmail.com>
Date: Fri, 05 Jul 2024 11:00:14 +0800
Message-ID: <000501dace87$75a315e0$60e941a0$@tsinghua.org.cn>
MIME-Version: 1.0
Content-Type: multipart/alternative; boundary="----=_NextPart_000_0006_01DACECA.83C76750"
Thread-Index: AQIR1ITPc6GYbMskLLmLageatLZx17F5XtqQ
Content-Language: zh-cn
Message-ID-Hash: LO4FMF4KR3KEL55FBAO7X7YEOJXJMX2W
Precedence: list
Subject: [Idr] 答复: [RTGWG & IDR] Discussing the Open Scheduled fabric of AI Network
Archived-At: <https://mailarchive.ietf.org/arch/msg/idr/DP-zEYDUv2C114UD9U9UsuoGep0>

Hi, Pengfei:

After reading your proposals, I have the following questions:

1. In your osf-framework, how you assure the sorting, fragmentation or aggregation of packet of one flow at the Ingress/Egress device does not lead to forwarding latency(also large required memory storage on these devices)?

2. For your congestion control proposal(credit-based flow control), what’s the difference from the traditional RSVP technology?

3. For your rapid link failure switchover in OSF, the main point is that the involvement of OSF-MS component that act as the central control point. Is it right?

Where is the OSF-MS component in your architecture that described in section https://datatracker.ietf.org/doc/html/draft-hcl-rtgwg-osf-framework-00#section-2.2

4. And, in https://datatracker.ietf.org/doc/html/draft-hcl-rtgwg-osf-framework-00#section-3, the document states that OSF-MS components can also operate in a distributed process mode, or a hybrid mode,

how do you achieve the rapid link failure switchover as AI applications required in these two modes?

Best Regards

Aijun Wang

China Telecom

发件人: forwardingalgorithm@ietf.org [mailto:forwardingalgorithm@ietf.org] 代表 霍朋飞
发送时间: 2024年6月30日 12:57
收件人: idr@ietf.org; rtgwg@ietf.org
主题: [Idr] [RTGWG & IDR] Discussing the Open Scheduled fabric of AI Network

Hi Wg,

Link:

<https://datatracker.ietf.org/doc/html/draft-hcl-rtgwg-ai-network-problem-00> https://datatracker.ietf.org/doc/html/draft-hcl-rtgwg-ai-network-problem-00

<https://datatracker.ietf.org/doc/html/draft-hcl-rtgwg-osf-framework-00> https://datatracker.ietf.org/doc/html/draft-hcl-rtgwg-osf-framework-00

<https://datatracker.ietf.org/doc/html/draft-hcl-idr-extend-tunnel-egress-point-00> https://datatracker.ietf.org/doc/html/draft-hcl-idr-extend-tunnel-egress-point-00

These documents describe the main disturbances caused by AI applications to the network and propose solutions based on open schedulable network architecture developed to address these issues.

It primarily focuses on the following aspects:

1. Open Scheduled Fabric

An open network architecture that resolves hash issues through dynamic scheduling based on network measurements.

2. Bgp Extension for Tunnel Egress Point

A method for implementing the OSF network architecture involves using BGP uniformly in the control plane, extending the data plane as a new tunnel encapsulation, and providing the necessary control information transmission for the data plane.

3. Extending the BGP protocol to carry Tunnel Egress Point information within the community attribute.

We welcome any comments and discussions.

Thank you!

Thanks,

Pengfei

[Idr] [RTGWG & IDR] Discussing the Open Scheduled… 霍朋飞
[Idr] 答复: [RTGWG & IDR] Discussing the Open Sched… Aijun Wang