Lima quotes a performance estimate: about 220 million rows across 615 tables, about 30 GB, in about 30 minutes on 64 vCPU. An estimate is a sizing calculation. This benchmark replaces it with a measured run anyone can repeat: a synthetic SAP planning dataset, a reference configuration, four scores and a published harness. The method is published now; the results follow next quarter; a terabyte run is scheduled after it.
Most performance figures in this category are quoted without the dataset, the hardware or the scoring. This page publishes those three first, so that when the results appear they can be checked, and so that anyone can run the same harness on their own tables and compare.
The figures Lima quotes today carry their labels: estimate, lab, production. The estimate is the one this benchmark exists to replace.
| Part | What it is | Why |
|---|---|---|
| Dataset | A synthetic SAP planning dataset generated with the shapes of real tables: material master (MARA, MARC), bills of material (MAST, STKO, STPO), routings (PLKO, PLPO), purchasing (EKKO, EKPO), sales (VBAK, VBAP), production (AFKO, AFPO), stock (MARD, MSEG). About 600 tables and 200 million rows, about 30 GB. | Real table shapes without real data; sized like a mid-sized estate. |
| Hardware | The 64C reference tier: one node, 64 vCPU, 512 GB, in Lima's cloud; the instance type is named in the harness. | The tier most estates start on. |
| Extraction methods | Each method timed on its own: table reads through the released function module, BAPIs, delta-enabled CDS views, change pointers, extractor delta queues; for Oracle, partitioned JDBC and LogMiner. | Rows per minute per method is the question buyers ask. |
| Transformation | The planning model for Kinaxis Maestro: Part, PartSource, BillOfMaterial, OnHand, ScheduledReceipt, IndependentDemand, Calendar. | The hardest shaping step in a planning load. |
| Load | Full load, then a net-change load of 2% of rows, into a Kinaxis test tenant in referential order. | Both the first load and the nightly case. |
| Scores | Rows per minute per method; hours to first full load; run-window adherence against a 4-hour window; reconciliation variance (rows and control totals). | The four numbers a planning program needs. |
| Harness | The generator, the pipeline definitions and the scoring script, downloadable when results publish; runnable on any Lima instance. | Reproducibility. |
A two-week bake-off runs the same protocol on your real SAP or Oracle tables and your planning system, and adds one confirmed write-back. The numbers are yours to keep, and they feed the field edition in anonymised form if you agree.
Because the run has to be made on the frozen method, and the method was frozen this month. Results follow next quarter, with the logs.
No. A generated dataset with realistic shapes, sized to terabytes. Customer runs stay on customer estates and enter the field edition only in anonymised aggregate.
Yes. The dataset generator and the scoring script are tool-neutral. Load the same dataset with any tool and score it the same way.
One email when the 64C results publish, one when the terabyte run does.
