[pim] Re: [Bier] Invitation to the "AI Network Multica st" side meeting at IETF 126

--===============5821604367969044929==
Content-Type: multipart/alternative; boundary="00000000000074afbe0656c9bd69"

--00000000000074afbe0656c9bd69
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

interesting enough, record pls if you can. overlaps with mpls so not a very
good slot to attend in person

thanks

-- tony

On Fri, Jul 17, 2026 at 9:36=E2=80=AFAM Yisong Liu <[email protected]=
om>
wrote:

> Dear All,
>
> We will organize a side meeting of AI Network Multicast on *Tuesday,
> July 21, 9:15-10:45* (Europe/Vienna), in *room Grand Klimt Hall 3*. If
> you are interested in AIDC or network layer multicast, please come to the
> side meeting to discuss. If you are remote participant, please access the
> meeting via the link: https://ietf.webex.com/meet/sidemeetings2
>
>
>
> This side meeting builds on previous side meeting discussions and will
> dive deeper into AI-specific multicast use cases and requirements,
> including MoE token dispatch, while conducting preliminary technical
> analyses of potential solutions such as BIER enhancements and new protoco=
l
> designs, with emphasis on reliability, scalability, and quantitative
> performance evaluation to guide future standardization.
>
>
>
> Please check the initial agenda of the side meeting in the following:
>
> *0. Welcome & Review  Chairs*
>
> *1. Multicast Problem Statement for AIDC  ( Junye Zhang, Remote, China
> Mobile)*
>
> *Abstract:* Multicast is a promising technique for enhancing the
> efficiency of point-to-multipoint data transmission during LLM training a=
nd
> inference in AI Data Centers (AIDCs). However, AIDCs demand bidirectional
> interaction, lossless reliability, microsecond scale dynamics, sparse
> groups, and simple control/data planes. This talk will give a brief probl=
em
> statement of multicast in AIDCs.
>
> *2.Can LLMs Reason Structurally? Benchmarking via the lens of Data
> Structures, with possible benefit from Multicast ( Yingxi Li, Remote,
> Stanford University)*
>
> *Abstract: *As large language models (LLMs) are increasingly deployed on
> complex tasks that demand multi-step decision-making, evaluating their
> algorithmic reasoning capabilities becomes critical. This talk introduces
> DSR-Bench, a diagnostic benchmark grounded in data structures, designed t=
o
> assess structural reasoning=E2=80=94the ability to perceive and manipulat=
e
> relational patterns such as order, hierarchy, and connectivity, including
> the benchmark design, the evaluation methodology, and key empirical
> findings. The talk concludes with a brief discussion of potential
> implications for AI data-center networks, raising the question of whether
> efficient multicast communication could benefit long-context inference,
> inference-time search, and distributed model execution.
>
> *3.In-Network Computing on Switches (IFEC): Remodeling Multicast and
> Accelerating Communication for AI Clusters  (Changwang Lin, On-site,
> Alibaba & H3C)*
>
> *Abstract: *Describe an innovative collaborative computing paradigm named
> IFEC (In Fabric Extended-Computation). Its core objective is to address t=
he
> worsening cross-xPU communication bottlenecks in large-scale AI training
> and inference, driven by rising model parallelism and the widespread
> adoption of architectures such as MoE.
>
> *4.  Multicast in Scale-Up (Haibo Wang, On-site, Huawei)*
>
> *Abstract:* With the rapid development of AI services, optimizing AI
> cluster performance has become increasingly important. Notably, some AI
> service scenarios may exhibit multicast-like communication patterns (e.g.=
,
> MOE dispatch).Traditional network multicast relies on single-direction
> forwarding, which conflicts with key AI service requirements:
> interactivity, reliability, dynamics, and sparsity. These mismatches
> introduce significant challenges in scale-out network. Interestingly, whe=
n
> AI services operate within a scale-up network, these challenges become mo=
re
> manageable. Consequently, we propose first focusing on leveraging multica=
st
> to enhance scale-up AI service performance, before extending the solution
> to broader distributed deployments.
>
> *5. Open Discussion*
>
>
>
> Welcome to join us! We hope to see you there!
>
> If you have any questions, please feel free to contact us:
> [email protected], [email protected]
>
> Best Regards
> Yisong
>
>
>
>
>
> _______________________________________________
> BIER mailing list -- [email protected]
> To unsubscribe send an email to [email protected]
>

--00000000000074afbe0656c9bd69
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr"><div>interesting enough, record pls if you can. overlaps w=
ith mpls so not a very good slot to attend in person=C2=A0</div><div><br></=
div><div>thanks=C2=A0</div><div><br></div><div>-- tony=C2=A0</div></div><br=
><div class=3D"gmail_quote gmail_quote_container"><div dir=3D"ltr" class=3D=
"gmail_attr">On Fri, Jul 17, 2026 at 9:36=E2=80=AFAM Yisong Liu &lt;<a href=
=3D"mailto:[email protected]">[email protected]</a>&gt; wro=
te:<br></div><blockquote class=3D"gmail_quote" style=3D"margin:0px 0px 0px =
0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div style=
=3D"background-color:rgb(255,255,255);color:rgb(0,0,0);line-height:1.5;word=
-break:break-all"><font face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91">Dear =
All,</font></div><div style=3D"background-color:rgb(255,255,255);color:rgb(=
0,0,0);line-height:1.5;word-break:break-all"><font face=3D"=E5=BE=AE=E8=BD=
=AF=E9=9B=85=E9=BB=91"><br></font></div><div style=3D"background-color:rgb(=
255,255,255);color:rgb(0,0,0);line-height:1.5;word-break:break-all"><p clas=
s=3D"MsoNormal"><font face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D=
"3">We will organize a side meeting of=C2=A0AI Network=C2=A0Multicast=C2=A0=
on=C2=A0<b>Tuesday, July=C2=A021, 9:15-10:45</b>=C2=A0(Europe/Vienna), in <=
b>room Grand Klimt Hall 3</b>. If you are interested in AIDC or network lay=
er multicast, please come to the side meeting to discuss. If you are remote=
 participant, please access the meeting via the link: <a href=3D"https://ie=
tf.webex.com/meet/sidemeetings2" target=3D"_blank">https://ietf.webex.com/m=
eet/sidemeetings2</a><u></u><u></u></font></p><p class=3D"MsoNormal"><font =
face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3">=C2=A0<u></u><u></=
u></font></p><p class=3D"MsoNormal"><font face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=
=85=E9=BB=91" size=3D"3">This side meeting builds on previous side meeting =
discussions and will dive deeper into AI-specific multicast use cases and r=
equirements, including MoE token dispatch, while conducting preliminary tec=
hnical analyses of potential solutions such as BIER enhancements and new pr=
otocol designs, with emphasis on reliability, scalability, and quantitative=
 performance evaluation to guide future standardization.<u></u><u></u></fon=
t></p><p class=3D"MsoNormal"><font face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=
=BB=91" size=3D"3">=C2=A0<u></u><u></u></font></p><p class=3D"MsoNormal"><f=
ont face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3">Please check t=
he initial=C2=A0<font>agenda of the side meeting in the following:=C2=A0</f=
ont><u></u><u></u></font></p><p class=3D"MsoNormal"><font face=3D"=E5=BE=AE=
=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3"><b>0.=C2=A0Welcome &amp; Review =C2=
=A0Chairs</b>=C2=A0<u></u><u></u></font></p><p class=3D"MsoNormal"><font fa=
ce=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3"><b>1.=C2=A0Multicast=
 Problem Statement for AIDC =C2=A0( Junye Zhang, Remote, China Mobile)</b>=
=C2=A0=C2=A0<u></u><u></u></font></p><p class=3D"MsoNormal"><font face=3D"=
=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3"><b>Abstract:</b> Multicast=
 is a promising technique for enhancing the efficiency of point-to-multipoi=
nt data transmission during LLM training and inference in AI Data Centers (=
AIDCs). However, AIDCs demand bidirectional interaction, lossless reliabili=
ty, microsecond scale dynamics, sparse groups, and simple control/data plan=
es. This talk will give a brief problem statement of multicast in AIDCs.<u>=
</u><u></u></font></p><p class=3D"MsoNormal"><font face=3D"=E5=BE=AE=E8=BD=
=AF=E9=9B=85=E9=BB=91" size=3D"3"><b><span>2.</span>Can LLMs Reason Structu=
rally? Benchmarking via the lens of Data Structures, with possible benefit =
from Multicast ( Yingxi Li, Remote, Stanford University)</b>=C2=A0<u></u><u=
></u></font></p><p class=3D"MsoNormal"><font face=3D"=E5=BE=AE=E8=BD=AF=E9=
=9B=85=E9=BB=91" size=3D"3"><b>Abstract: </b>As large language models (LLMs=
) are increasingly deployed on complex tasks that demand multi-step decisio=
n-making, evaluating their algorithmic reasoning capabilities becomes criti=
cal. This talk introduces DSR-Bench, a diagnostic benchmark grounded in dat=
a structures, designed to assess structural reasoning=E2=80=94the ability t=
o perceive and manipulate relational patterns such as order, hierarchy, and=
 connectivity, including the benchmark design, the evaluation methodology, =
and key empirical findings. The talk concludes with a brief discussion of p=
otential implications for AI data-center networks, raising the question of =
whether efficient multicast communication could benefit long-context infere=
nce, inference-time search, and distributed model execution.<u></u><u></u><=
/font></p><p class=3D"MsoNormal" style=3D"margin-left:0pt;text-indent:0pt">=
<font face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3"><b><span>3.<=
/span>In-Network Computing on Switches (IFEC): Remodeling Multicast and Acc=
elerating Communication for AI Clusters =C2=A0(Changwang Lin, On-site, Alib=
aba &amp; H3C)</b><u></u><u></u></font></p><p class=3D"MsoNormal"><font fac=
e=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3"><b>Abstract: </b>Desc=
ribe an innovative collaborative computing paradigm named IFEC (In Fabric E=
xtended-Computation). Its core objective is to address the worsening cross-=
xPU communication bottlenecks in large-scale AI training and inference, dri=
ven by rising model parallelism and the widespread adoption of architecture=
s such as MoE.<u></u><u></u></font></p><p class=3D"MsoNormal"><font face=3D=
"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3"><b>4.=C2=A0=C2=A0Multicas=
t in Scale-Up (Haibo Wang, On-site, Huawei)</b><u></u><u></u></font></p><p =
class=3D"MsoNormal"><font face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" siz=
e=3D"3"><b>Abstract:</b> With the rapid development of AI services, optimiz=
ing AI cluster performance has become increasingly important. Notably, some=
 AI service scenarios may exhibit multicast-like communication patterns (e.=
g., MOE dispatch).Traditional network multicast relies on single-direction =
forwarding, which conflicts with key AI service requirements: interactivity=
, reliability, dynamics, and sparsity. These mismatches introduce significa=
nt challenges in scale-out network. Interestingly, when AI services operate=
 within a scale-up network, these challenges become more manageable. Conseq=
uently, we propose first focusing on leveraging multicast to enhance scale-=
up AI service performance, before extending the solution to broader distrib=
uted deployments.</font></p><p class=3D"MsoNormal"><font face=3D"=E5=BE=AE=
=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3"><b>5.=C2=A0Open Discussion</b><u></=
u><u></u></font></p><p class=3D"MsoNormal"><font face=3D"=E5=BE=AE=E8=BD=AF=
=E9=9B=85=E9=BB=91" size=3D"3">=C2=A0<u></u><u></u></font></p><p style=3D"m=
argin-top:0pt;margin-left:0pt;text-indent:0pt;padding:0pt;background:rgb(25=
5,255,255)"><font face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=3D"3">=
Welcome to join us! We hope to see you there!<u></u><u></u></font></p><p cl=
ass=3D"MsoNormal"><font face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91" size=
=3D"3">If you have any questions, please feel free to contact us: <span><a =
href=3D"mailto:[email protected]" target=3D"_blank">liuyisong@china=
mobile.com</a></span>, <span><a href=3D"mailto:[email protected]" target=
=3D"_blank">[email protected]</a></span></font></p></div><div style=3D"ba=
ckground-color:rgb(255,255,255);color:rgb(0,0,0);line-height:1.5;word-break=
:break-all"><br></div><div style=3D"background-color:rgb(255,255,255);color=
:rgb(0,0,0);line-height:1.5;word-break:break-all"><font face=3D"=E5=BE=AE=
=E8=BD=AF=E9=9B=85=E9=BB=91">Best Regards</font></div><div style=3D"backgro=
und-color:rgb(255,255,255);color:rgb(0,0,0);line-height:1.5;word-break:brea=
k-all"><font face=3D"=E5=BE=AE=E8=BD=AF=E9=9B=85=E9=BB=91">Yisong<br></font=
><br><br><br><br><font face=3D"=E5=AE=8B=E4=BD=93" style=3D"font-size:16px"=
>=C2=A0</font></div>_______________________________________________<br>
BIER mailing list -- <a href=3D"mailto:[email protected]" target=3D"_blank">bie=
[email protected]</a><br>
To unsubscribe send an email to <a href=3D"mailto:[email protected]" targ=
et=3D"_blank">[email protected]</a><br>
</blockquote></div>

--00000000000074afbe0656c9bd69--


--===============5821604367969044929==
Content-Type: text/plain; charset="utf-8"
MIME-Version: 1.0
Content-Transfer-Encoding: base64
Content-Disposition: inline

X19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX19fX18KcGltIG1haWxp
bmcgbGlzdCAtLSBwaW1AaWV0Zi5vcmcKVG8gdW5zdWJzY3JpYmUgc2VuZCBhbiBlbWFpbCB0byBw
aW0tbGVhdmVAaWV0Zi5vcmcK

--===============5821604367969044929==--