DASH Workgroup Meeting Minutes 3/6/2024
Kristina Moore
Hello DASH Community –thank you for your time last Wednesday.
We had @Srikanth Kandula from Microsoft Research present In Network Transformations for AI/ML and other Workloads. This was an interesting and informative talk around development of an open API for in-network transformations, and early results. The slides have been uploaded to the DASH GitHub if you would like to view them.
Intel @'Cristian Dumitrescu' provided additional clarification around ‘range-match’ and the limitations of range support in P4 (rather than bmv2).
We were also able to close
5 PRs/Issues this week (see project list).
For further reading, see full meeting notes and follow-ups below.
Thank you for your time/contributions ~ Kristina
As always, with respect to
any of our DASH meetings - Community (Wednesdays), or Behavioral Model (Thursdays) – please let me know if there are PRs, Q&A, or items you would like to discuss or present.
The DASH YouTube channel link is here to
subscribe / access WG content (and click the bell to receive notifications).
See you on 3/13/2023!
Meeting Title: SONiC-DASH-Workgroup Community Meeting #104
Attendees (20):
DASH Group to join: https://groups.google.com/g/sonic-dash
DASH-Test-Workgroup Group to join: https://groups.google.com/g/sonic-dash-test-workgroup
|
Agrawal, Ashutosh - Intel |
Guohan Lu - MSFT |
murali Venkateshaiah - Cisco |
Rob Murphy - Cisco |
|
Alberto Villarreal - Keysight |
Kristina Moore - MSFT |
mxiao - Arista |
Srikanth Kandula - MSFT |
|
Chris Sommers - Keysight |
Kumaresh Perumal - MSFT |
Oleksandr Ivantsiv - NVidia |
Sushant Sharma - MSFT |
|
Don Ewald - Cisco |
Mididaddi, Naren - Intel |
Praveen Bhagwatula - Cisco |
Veerappan, Senthilnathan - AMD |
|
Dumitrescu, Cristian - Intel |
Mircea Dan Gheorghe - Keysight |
Riff Jiang - MSFT |
Yakiv Huryk - NVidia |
Discussion - PRs/Issues/Documentation for review, comments, suggestions
DASH Community Upcoming Project Action Items
TL;DR Notes 😊
- Christian (Intel) clarified the limitations of P4 DPDK for range support with the Community; issue is not due to bmv2, rather P4.
-
In Network Transformations for AI/ML and other Workloads: In-network compute DASH API - Srikanth Kandula, Partner Researcher at Microsoft Research
https://www.linkedin.com/in/srikanthkandula/
- Early joint work with Rishabh, Deepak, Gerald and Srikanth
- Implementations clean & clear
- Providers able to deploy different implementations
- App libraries to have a consistent coding interface
- Srikanth introduced the motivation and objectives of developing an open API for in-network transformations for AI/ML workloads, such as multicast, aggregation and compression, and discussed the design choices and challenges with the community.
- Feedback and questions for Srikanth: the Community raised several questions and comments for Srikanth, such as the scalability and memory requirements of the transformations, the coexistence with different network protocols and stacks, the failover and load balancing mechanisms, and the data plane signaling and identification.
- Q: Robert Murphy (Cisco) - scale needed for distribution trees? Relates to memory consumption…
- A: Srikanth (MSR) - roughly 1 tree for ongoing collective. Depends on the job and how long it spends time in the GPU or collective. # or training jobs or inferences running in the cluster.
Follow-up tasks:
- Network compute API: Obtain information on how deep in the packet the NPU can look for the signaling field (@robe...@cisco.com)
- Network compute API: Send the meeting series to Srikanth and forward any feedback from the community (Kristina)
- Upload presentation slides to GitHub - done
Each receiver needs to receive other 3 packets; multicast protects the out-link
Math looks like matrix multiplier for an outcome of a smaller vector
Key takeaway is that good results are being seen
Think of these as axioms
Device doing this needs to maintain the state
Behavioral Model WG - March 7 2024 - https://youtu.be/DveEx4nxgug
HA APIs from Riff from a SAI perspective to support DASH HA and ENI level HA
Riff presented the high-level concepts, data structures, packet flow and workflows for the DASH HA API design and answered questions from Intel, and Keysight. The API is intended to support different HA modes and scenarios and is being vetted by some backchannel stakeholders.
DASH HA P4 implementation: Riff also shared his progress on the P4 implementation of the DASH HA behavior model and the plan to create a virtual DPU image for testing. He mentioned some challenges with the data plane app and the flow API. 35:39
P4-DPDK for Dash: Chris Sommers updated the group on Andy's work on P4-DPDK for DASH and the issues he encountered with P4 runtime. He said the P4-DPDK team is aware of the issues filed with the P4-DPDK team related to the DASH use case. 43:41
Other Notes:
Introduction/Welcome:
Srikanth Kandula, Partner Researcher at Microsoft Research
https://www.linkedin.com/in/srikanthkandula/
|
Sticky for Links/Data: |
|
|
|
DASH Groups to join to receive Invites, Meeting Notes, and Comms |
DASH: https://groups.google.com/g/sonic-dash
DASH-Test-Workgroup Group: https://groups.google.com/g/sonic-dash-test-workgroup |
If anyone knows other people who would like info re: our community, please have them joins these groups for receive Comms, etc… |
|
Links to Recording |
Teams/Sharepoint:
DASH Community YouTube:
|
3/6/2024 Community Call; please request access via the link if you are not able to view/listen
|
|
Azure DASH GitHub Repo: |
|
|
|
|
https://github.com/Azure/DASH/blob/main/test/docs/dash-test-workflow-saithrift.md |
Ideal test workflow is here, converted to .md |
|
SAI Thrift |
Client server needed for testing |
|
|
P4 |
https://opennetworking.org/p4/ and https://p4.org/working-groups/ |
Open source, domain-specific programming language for network devices, specifying packet processing for data plane devices (switches, routers, NICs, filters, etc.) |
|
PINS |
|
|
|
PNA consortium spec |
An architecture describing the structure and common capabilities of network interface controller (NIC) devices which process packets transiting one or more interfaces
and a host system. |
|
|
IPDK |
IPDK is an open source, vendor agnostic framework of drivers and APIs for infrastructure offload and management which runs on a CPU, IPU, DPU or switch. IPDK runs in Linux and uses a set of well-established tools such as DPDK and P4 to enable network virtualization. |
|
|
bmv2 |
The second version of the reference P4 software switch, nicknamed bmv2 (for behavioral model version 2). The software switch is written in C++11. It takes as input a JSON file generated from your P4 program by a P4 compiler and interprets it to implement the packet-processing behavior specified by that P4 program |
|
|
DPDK |
DPDK is the Data Plane Development Kit which consists of libraries to accelerate packet processing workloads running on a wide variety of CPU architectures. |
|
|
Linux Foundation SmartSwitch |
https://lists.sonicfoundation.dev/g/sonic-smartswitch/calendar |
|
Thank you again for your participation…
Kristina Moore MBA,
M.S., CISSP - Azure Core Principal PM / DASH
Office: 425-722-7720 Mobile: 425-876-2040 Email:
kri...@microsoft.com
DASH Group to join:
https://groups.google.com/g/sonic-dash
DASH-Test-Workgroup Group to join: https://groups.google.com/g/sonic-dash-test-workgroup
