S4: Computational storage
Question: Can we verify an offloaded computation?
Time: 30 minutes. Prerequisites: completed setup. See qualification and limitations before running.
Mechanism
- Allocate device memory and transfer a deterministic input.
- Load a trusted host program and execute it through the CSD interface.
- Retrieve the result and compare it with the CPU reference.
Predict
What would a successful command tell us if the returned data were wrong?
Write down a predicted state or counter change before executing the commands.
Run
Host terminal, unless a command is explicitly labelled Guest. Source and
export the identity fields from connection.env. These labs overwrite the
explicit disposable namespace. Use a fresh VM for each independent arm.
# Provision csd-passthru and csd-vadd.so from the pinned source.
bash labs/bin/preflight.sh labs/guest/lab-s4-csd.sh /dev/nvme0n1
Check your evidence
- The real vector-add result equals the reference output.
- Allocation, load, activation, execution and cleanup all succeed.
- Program and tool hashes accompany the transcript.
The host runner retrieves raw receipts and preserves the guest exit status. Keep unsuccessful attempts; do not replace them with a later successful log.
Interpret
What data movement should a useful offload eliminate?
Explain the mechanism and the measurement boundary. A throughput number by itself is not evidence that the intended feature was exercised.
Recover
- Program missing: inspect the host csd_program_dir.
- Unknown command: rebuild the guest tool from this source revision.
- A model runtime is not an independent hardware speedup measurement.
Save the receipt before resetting the VM.
Materials and source
Slides, scripts and worksheet.
Videos are not yet recorded. The frozen public source is
1918eb9795ba.
The FEMU Manual explains the implementation and supported properties.