In 2005 the Xen team asked whether VMMs are microkernels done right; over twenty years isolation went to microVMs, the dataplane to silicon, discipline to DPUs, and only 'the service stack restarts painlessly' went unclaimed — this article traces that history and its four obstacles, then offers a pure-software assembly: a thin core defined as firmware, stateless service VMs, a dataplane delegated to silicon, holding with VM boundaries and the IOMMU the boundary the DPU holds with PCIe.
Every large-scale DPU success comes from a hyperscaler's vertically integrated, in-house program, while the merchant DPU market is a complete wreck. An attribution audit of its seven value propositions: five do not survive scrutiny, one is real but software-replaceable, and one is real but worth only half — half a claim left standing out of seven. Thin value wrapped in a thick narrative is the standard recipe for a myth.
Fixed-output compression — variable-length input stuffed into fixed-size output blocks — was credited with eliminating read amplification, aligning I/O, improving compression ratio, and saving memory. Term-by-term attribution finds otherwise: these benefits belong to granularity choices, to the index and the decompressor, hold only inside a restricted framing, or cannot be found in measurement at all — while the costs are format-level and permanent. The conclusion is polite but total: as a format design decision, fixed-output is not worth adopting.
OCI tar layers for containers, qcow2 for VMs — both carry structural costs. Overlaybd is a layered, lazily-loaded, seekably-compressed block-device image format for every workload: sub-second cold starts, proven in production across many companies.
The first post of this blog