Every layer of the storage stack. One specialty. Storage.
From raw NVMe commands to the GPU's data path, our engineers have shipped production code at each layer. When storage is the bottleneck, we can follow the problem through every layer ourselves.
- Built production code our engineers wrote
- Upstream features in the open-source project's own codebase
- Deployed & tuned installed, benchmarked, handed over
↑ application / GPU
Workloads HPC · AI · Kubernetes
- Go CSI driver serving block volumes to containers and KubeVirt VMs (Built)
- Slurm scheduling and xCAT provisioning for HPC clusters (Deployed & tuned)
Observability eBPF
- Fleet-scale eBPF tracing we built (Built)
Caching & tiering KV cache · cold tier
- S3 garbage collector keeping the cold tier consistent (Built)
- LMCache KV-cache tiers for LLM inference (Deployed & tuned)
Filesystems & protocols Lustre · Ceph · pNFS
- BeeGFS core development with ThinkParQ: metadata search, mirroring, CLI (Upstream)
- Lustre, Ceph and pNFS deployment and tuning (Deployed & tuned)
Control plane & consensus Raft · RocksDB
- Raft-replicated, RocksDB-backed metadata service with leases (Built)
Data servers & protection replication · EC · recovery
- Chunk replication, erasure coding, background GC and recovery (Built)
Kernel, block & devices io_uring · ublk · NVMe
- ublk block driver (DKMS), io_uring reactor, raw NVMe passthrough (Built)
- Filesystem client compatibility through Linux 6.12 (Upstream)
↓ NVMe / devices
Instruments
We build the tools when the standard ones run out
We show exactly which kernel function, I/O path and node is the bottleneck, without redeploying, recompiling or instrumenting your application. Then we build eBPF tools so your team can find the next bottleneck.
trace.yaml · what to trace
target: vfs_read measure: latency output: histogram nodes: storage-*
vfs_read.bt · generated (illustrative)
kprobe:vfs_read /comm != "bpftrace"/ {
@start[tid] = nsecs;
}
kretprobe:vfs_read /@start[tid]/ {
@latency_us = hist((nsecs - @start[tid]) / 1000);
delete(@start[tid]);
}A runnable bpftrace script, with kprobe, uprobe or tracepoint attachment chosen automatically, scheduled across the fleet.
Custom built
yapio
A parallel I/O benchmark we wrote in C, where groups of processes run different I/O patterns across nodes at the same time.
Custom built
eBPF tracing
Follows slow I/O from application calls through the kernel to the device.
Run at scale
The standard toolchain
- IOR / mdtest
- fio
- IO500
- TLA+ / TLC
- bpftrace
- Ansible
- pytest
- Grafana
Failure testing
Proven in the lab, ready for production
Distributed systems fail in ways that only real clusters reveal. We build deterministic automation that stands up real multi-node clusters, injects exact failures, and gates every release on the result.
Node crash
Clients reconnect, no acknowledged write is lost
● encoded in tests
Network partition
Consensus holds, no split-brain, clean rejoin
● encoded in tests
Disk failure
Chunk recovery completes, every checksum matches
● encoded in tests
Failure mid-resync
Mirrored metadata survives a second failure
● encoded in tests
Live upgrade
Two major versions mount side by side; rollback is safe
● encoded in tests
Pod restart
CSI block devices keep serving in-flight I/O
● encoded in tests
Find the layer that's costing you throughput.
Send us your topology, workload and a benchmark run. We will get back to you.