Design distributed systems that hold up from day one
We plan the cluster and storage layout, and model the software formally in TLA+ to prove it stays correct, even in failure scenarios, before anything is built.
What we offer
Cluster architecture
We turn your workloads, hardware and capacity targets into a reference architecture, sizing and capacity plans.
Storage architecture
Chunk placement, replication and erasure coding, highly available metadata, garbage collection and cold tiering.
Formal software modelling
Model distributed system software in TLA+ and check it with TLC, so protocol and consensus bugs are caught before any code is written.
Expertise
From NVMe to GPU. Every layer covered.
Scalable storage is only as fast as its slowest layer. Pick a discipline to see where in the stack we work.
- HPC, AI & Kubernetes workloadsMPI · PyTorch · CSI
- ObservabilityeBPF · bpftrace
- Caching & tieringLMCache · S3 cold tier · GC
- Filesystems & protocolsLustre · Ceph · pNFS
- Control plane & consensusRaft · RocksDB · leases
- Data servers & protectionreplication · EC · recovery
- Kernel, block & devicesio_uring · ublk · NVMe · ZFS
Kernel modules and storage engines that stay fast and correct across new Linux kernel versions: reactors, custom allocators, raw NVMe commands.
Highly available metadata services on a Raft-replicated, RocksDB-backed store, with leader election, lease coordination and failure detection.
Building new filesystem features, our deepest expertise, plus deployment and tuning of Lustre, Ceph and pNFS.
Block storage delivered natively to Kubernetes, for containers and KubeVirt VMs alike, through a CSI driver we implemented, with an S3 garbage collector that keeps tiered data consistent and reclaims space in the object tier.
We build block storage from the ground up: chunk placement, replication and erasure coding, garbage collection, cold tiering and native Kubernetes volumes.
We design the cluster before anyone racks it, with reference architectures, hardware sizing and capacity plans, then make it highly available with DRBD, provision nodes at scale with xCAT and integrate Slurm scheduling.
Connecting AI frameworks to high-performance storage so GPUs compute instead of waiting, from LMCache KV-cache tiers to direct NVMe-to-GPU paths.
Kernel-level tracing of block I/O, VFS and function latency across a fleet, with zero application instrumentation.
Deterministic multi-node test harnesses, formal verification and performance pipelines that measure throughput and recovery together.
Storage slower than your compute?
Send us your topology, workload and a benchmark run. We will get back to you.