035
What Live Optics collects, and what it quietly does not
First in a short series on the tool everyone runs before a storage or platform proposal. Knowing the shape of the data is the difference between a sizing and a guess.
Articles
Everything published so far, newest first.
034
Twenty-four hours of Live Optics is not a workload profile
Second in the series. The collection window is the single input that most often invalidates a sizing, and it is the one nobody negotiates.
033
Reading Live Optics storage numbers without being led
Third in the series. IOPS, block size, read/write mix and the compression estimate — what each one is actually measuring, and where the sizing goes wrong.
032
From collection to proposal: write the assumptions on the page
Last in the Live Optics series. The arithmetic between the data and the recommendation is where every argument happens, and it is almost always invisible.
031
The order I read a Broadcom renewal quote in
Most people open the VMware renewal at the total. That number is the last thing you should look at, and by then you have already lost the argument.
030
What a parallel file system actually is
It is not a faster NAS. The architectural difference is that clients talk to many storage servers at once, and almost every operational surprise follows from that.
029
Metadata is the bottleneck, not bandwidth
Every parallel file system is sold on GB/s. Almost every one that disappoints in production is disappointing because of operations per second on small files.
028
Stripe count, stripe size, and the defaults that cost you most of your bandwidth
Two numbers control whether your parallel file system behaves like dozens of servers or like one. Most sites never change them.
027
Lustre in the real world: what operators actually complain about
The performance is not the issue. The issues are upgrades, client evictions, one slow target poisoning everything, and the fact that it assumes you employ someone who knows it.
026
IBM Storage Scale: what it does well, and the licensing conversation
GPFS by its older name. Technically excellent, operationally more forgiving than Lustre, and sold in a way that requires you to read carefully.
025
BeeGFS and the cluster that is too big for NFS and too small for Lustre
There is a real gap in the middle of this market, and the thing that fills it is usually the one nobody shortlisted.
024
All-flash parallel file systems: what changed and what did not
Flash removed the seek penalty that shaped twenty years of design. It did not remove the metadata bottleneck, the network limit, or the need to think about layout.
023
Sizing a parallel file system: start with bandwidth, not terabytes
Capacity is the easy number and the wrong one to start from. Here is the order that produces a design somebody can defend.
022
Checkpointing is the write burst that sizes your file system
A cluster that reads steadily and writes rarely still needs to absorb its entire memory footprint in a few minutes. That burst, not the average, decides the design.
021
Nutanix is not a drop-in for vSphere, and pretending otherwise costs you month four
AHV will run your VMs. That was never the risky part. The risk is in the twenty things around the VM that you forgot were vSphere features.
020
GPUDirect Storage: what it buys you, and when it does not
Moving data from storage into accelerator memory without a detour through the host. Real gains in specific conditions, and a long list of prerequisites.
019
The small-file problem, and how machine learning made it worse
Every shared file system struggles with millions of tiny files. Training pipelines produce exactly that, and the fix is in the pipeline rather than the storage.
018
How parallel file systems fail: the modes nobody warns you about
They rarely stop. They degrade, partially and confusingly, in ways that present to users as 'the storage is slow' with no pattern.
017
You cannot back up a parallel file system the way you think
Walking a namespace of a billion files with a conventional backup agent will not finish. The strategies that work look nothing like enterprise backup.
016
Sizing a backup appliance: the five numbers that matter
Front-end capacity, change rate, retention, ingest rate and restore rate. Get these and the configuration follows. Skip one and the appliance is wrong.
015
Deduplication ratios are a marketing number until you name the dataset
A 50:1 ratio and a 4:1 ratio can both be honest. The difference is what was measured, and the assumption in your quote determines whether the appliance fits.
014
Nobody sizes for restore
Backup windows get modelled to the hour. The restore rate, which is the number that matters on the day something has gone wrong, is usually nowhere in the document.
013
The backup appliance market: who is winning, and at what
The category split in two. Dedicated backup storage is consolidating around fewer hardware vendors while the growth moved to software platforms that sell cyber resilience.
012
Parallel file system or scale-out NAS: the decision most teams get backwards
The default answer for an enterprise AI cluster should be NAS, and the burden of proof should sit with the parallel file system. Usually it is the other way round.
011
Inference is not training, and your sizing model probably assumes it is
Two workloads with opposite characteristics, sized with the same reference architecture. The result is a cluster that is wrong in both directions at once.
010
What OpenShift Virtualization quietly asks of your storage team
KubeVirt runs VMs inside pods. That sentence hides an entire storage conversation that most migration plans discover in the pilot.
009
Object storage for AI datasets: cheaper, slower, and often the right answer
Not everything needs to sit on the fast tier. The interesting design question is what moves between tiers and who decides.
008
An RVTools export is not an assessment
Everybody starts with the same spreadsheet. Almost nobody reads the twelve columns that change the answer, and every vendor tool quietly fills the gaps in its own favour.
007
What a GPU node costs before it produces a single token
The accelerators are the headline. Power, cooling, fabric, storage and the eighteen weeks of waiting are the budget.
006
Nobody sizes the fabric, then blames the GPUs
East-west traffic in an AI cluster behaves nothing like the enterprise traffic your network was designed for. The symptom looks like slow accelerators.
005
Indexing 604 pages of recitation with Whisper and a lot of fuzzy matching
A weekend project that taught me more about applied AI than any vendor deck: transcription is the easy half, alignment is the real problem, and the human in the loop is not a failure.
004
RAG is a retrieval problem wearing an AI costume
Nine out of ten disappointing RAG systems are disappointing search systems. The generation step is rarely where the failure is.
003
NVMe/TCP quietly won the mid-range while everyone argued about Fibre Channel
Not because it is faster. Because it removed a team, a skill set and a second network from the build sheet.
002
Your immutable backup is not an air gap
Both words get used as though they mean the same protection. They do not, and the difference is the credential that can reach the repository.
001
The three questions that end most 'we will just move to the cloud' plans
Not arguments against public cloud. Just the three numbers that tend to be missing when the business case looks obviously good.