Run Control
Runs are limited to the two published research examples. Each result identifies whether it was generated now or captured earlier.
STATEFUL / P4
00 / Transparent Research Build
SpliDT
...
Loading
Evidence Status
Loading available sources...
Current Run
Ready
Choose what to inspect, then start.
01 / Packet Journey
PACKET TO DIGEST
Watch ten packets form two feature windows. At each window boundary, the active subtree evaluates real if/else thresholds, converts the leaf class into the next SID, and either persists that state or emits the final digest.
Partitioned execution
02 / Execution Trace
EVERY STEP
WAITING TO START
Select a model and source above. The resulting stages, commands, inputs, outputs, and logs will appear here.
03 / Proof Files
RAW EVIDENCE
WAITING FOR A RESULT
Model files, mappings, P4 output, target state, timings, and checksums will appear here.
04 / Research Record
FROM PROPOSAL TO PIPELINE
Open source research / 2025
SpliDT was developed for the P4 Language Consortium during Google Summer of Code 2025.
The public showcase connects the original partitioned-tree idea to generated P4, control-plane rules, packet state, and verifiable target evidence.
June 2025 / KickoffThe proposed compiler-to-target pipeline
September 2025 / Wrap-upThe implemented compiler, control plane, and data plane
05 / Systems In The Loop
THE ACTUAL STACK
ACADEMIC ROOTS. INDUSTRIAL TARGETS. OPEN GOVERNANCE.
P4 grew from the programmable-networking research community around Stanford and its collaborators. Today it is a global open-source project hosted by the Linux Foundation; SpliDT applies that lineage to stateful ML inference on software and Intel Tofino targets.
01 / DATA PLANE
P4 + P4Runtime
Expresses the packet pipeline and exposes its tables, registers, digests, and control-plane contract.
02 / ASIC TARGET
Intel Tofino + TNA
Provides the hardware architecture targeted through Open-P4Studio and the recorded Tofino model evidence.
03 / CONTROL CHANNEL
gRPC
Carries P4Runtime writes, register interactions, rule installation, and digest communication.
04 / SOFTWARE TARGET
BMv2 + Mininet
Compiles and exercises the generated pipeline on Linux before hardware-target validation.
05 / GENERATION
Python + Jinja2
Transforms partitioned trees into P4 source, table entries, SID mappings, and runtime control code.
06 / MODEL SEARCH
TensorFlow + scikit-learn
Supports training and evaluation around the HyperMapper-guided design-space exploration loop.
06 / People Behind SpliDT
THE RESEARCH NETWORK
GSoC 2025 contributor / implementation lead
SANKALP JHA
Turned the research design into an inspectable P4 implementation: tree conversion, SID-based traversal, generated data-plane code, runtime integration, target experiments, and this public evidence-driven showcase.
Portfolio / Work log
Primary mentors
DIRECT TECHNICAL GUIDANCE
University of Michigan / SpliDT co-authorAnnus Zulfiqar
Primary mentor for the 2025 implementation and a core researcher behind SpliDT's programmable-system design.
University of Michigan / P4 systemsAli ImranPrimary mentor for the implementation, bridging the research architecture with practical P4 and runtime engineering.
Supporting mentors
RESEARCH + P4 ECOSYSTEM
Nokia / P4 practitionerDavide Scano
Supported target-facing P4 work with experience across Tofino, telemetry, and programmable-network deployments.
NIKSUN / SpliDT co-authorWalter WillingerConnected the implementation to the wider networking-research perspective and SpliDT's systems evaluation.
University of Michigan / SpliDT co-authorMuhammad ShahbazSupported the project through expertise in programmable networks, data-plane systems, and in-network ML.
Purdue University / SpliDT first authorMurayyiam ParvezHelped carry the original partitioned-inference research into the GSoC implementation and evaluation path.
The broader SpliDT paper team spans Purdue University, the University of Michigan, UC Santa Barbara, the Open University of Israel, and NIKSUN. Mentor roles above follow the official P4 GSoC 2025 project listing.
07 / Reading Guide