Performance software
Validate your claims without renting a cluster.
When you would use this.
The same measurement answers a different question at each point.
Validate your claim during R&D
Find out how your product performs during development, before you commit to a number, using workloads, hardware, and topologies at the scale you need.
Demonstrate your claim to customers
Show customers how your product performs at the scale they care about, using their workloads and topologies.
Keep the claim current
New accelerators, new fabrics, new model architectures. A number measured on last generation’s hardware says nothing about this one, so re-run it and publish a figure that still holds.
Your product, someone else’s workload.
A gain measured on a workload you chose is the benchmark buyers already discount. This runs against traffic a real model produces.
Where it comes from
Scala’s library of captured workloads, or a trace from the customer you are trying to win, with their agreement. Either way the communication pattern is one a real model produced rather than one written to suit your product.
What gets compared
The same workload on the same modeled cluster, with and without your product. Change the scale and run it again to find where the gain grows, holds, or disappears.
Sweeps run through the API, so a range of scales is one job rather than a series of engagements.
Review the performance claim you need to prove.
Twenty minutes on the claim you need to stand behind, what we would need from you, and what the comparison would look like.