This paper [1] is published in ICDCS 2024. Its title "AdapCC: Making
Collective Communication in Distributed Machine Learning Adaptive". It is
relevant to the discussion of this sidemeeting. You may think of inviting
the authors to present this paper in future Inter-DC AI sidemeeting.
Hesham
[1] https://i.cs.hku.hk/~cwu/papers/xyzhao-icdcs24.pdf
On Fri, Jul 19, 2024, 2:55 AM Dirk Trossen wrote:
> Dear all,
>
>
>
> We are organizing a side meeting on “Inter-DC AI: Requirements and
> Challenges” on Thursday 25th from 5pm to 6.15pm in Prince of Wales/Oxford.
>
>
>
> The backdrop for this side meeting is the significant growth in demands
> for computing and networking resources for (Large data) AI, where
> boundaries for future growth are being set through power consumption
> (including location of the DC), space and cooling tech as well as
> complexity and cost. This has led to the emerging view that inter-DC AI
> computation is the way forward to overcome those local DC challenges.
>
>
>
> This side meeting at IETF120 will start the needed dialogue to identify
> and discuss key challenges on (1) congestion control to improve goodput
> across DCs, (2) efficient support for cross-DC collective communication
> primitives, (3) obtaining suitable knowledge of network topologies across
> underlays to segregate training tasks across DCs, (4) utilizing new
> congestion control mechanisms to optimize large-scale inferencing traffic
> from millions of clients to few Points of Presence (PoPs), as well as (5)
> providing suitable means to attest for secure and private transfer of
> needed training input data from customers to AI training providers.
>
>
>
> The side meeting will feature presentations to tease out key insights from
> various players in this field, leaving dedicated time for discussing
> possible next and concrete steps in the IETF to move forward, with the
> following agenda
>
>
>
> Room: Prince of Wales/Oxford Agenda: 17.00 - 17.05: Settling &
> Introduction (Luigi Iannone, Huawei)
>
>
>
> 17.05 - 17.15: Inter-DC AI: Requirements & Opportunities (Dirk Trossen,
> Huawei)
>
>
>
> 17.15 - 17.25: On Congestion Control (Michael Welzl)
>
>
>
> 17.25 - 17.35: On Attestation (Ramki Krishan, Intel)
>
>
>
> 17.35 - 17.45: On Collective Communication (Kehan Yao, China Mobile)
>
>
>
> 17.45 - 18.15: Discussion on Next Steps (Dirk Trossen, Huawei)
>
>
>
> Meeting link: https://ietf.webex.com/meet/ietfsidemeeting2
>
> Github link (with presentation slides coming next week):
> https://github.com/dirk-trossen-huawei/ietf120_inter-dc_ai
>
>
>
> We hope to see you on 25th for a fruitful and engaging discussion!
>
>
>
> Best,
>
>
>
>
>
> Dirk Trossen, Luigi Iannone, David Lou
>
>
> ___
> rtgwg mailing list -- rtgwg@ietf.org
> To unsubscribe send an email to rtgwg-le...@ietf.org
>
___
rtgwg mailing list -- rtgwg@ietf.org
To unsubscribe send an email to rtgwg-le...@ietf.org
