What is theoretical maximum throughput model for single Exadata rack?
To model the theoretical maximum throughput of a single Exadata X11M (2026-era) rack, we must move beyond "Marketing IOPS" and look at the physical limits of the PCIe Gen5/6 bus, the DDR5 memory controllers, and the 100/200GbE RoCE fabric.
In a 1M TPS architecture, the "Maximum" isn't a single number—it is a balanced equation where the bottleneck shifts depending on whether the workload is CPU-bound (OLTP) or I/O-bound (Analytics).
To build the model, we calculate the "Ceiling" for each hardware layer in a full rack (typically 8 Database Servers and 14 Storage Cells).
Exadata uses a redundant 100Gbps (or 200Gbps in X11M) RoCE Network Fabric.
Theoretical Throughput: With 14 Storage Cells providing data to 8 DB Nodes.
The Math: $14 \text{ cells} \times 200\text{Gbps} \approx 2.8\text{ Terabits/sec}$ of raw fabric bandwidth.
The Reality: After protocol overhead (iDB/RDMA), the fabric supports $\sim 250\text{--}300\text{ GB/s}$ of sustained data movement.
Each Storage Cell contains multiple NVMe Gen5 drives.
The Math: $14\text{ cells} \times 4\text{ NVMe drives} \times 14\text{ GB/s (Gen5 peak)} \approx 784\text{ GB/s}$.
The Constraint: The Flash is faster than the Network. This means in a full-rack "Smart Scan," the Network Fabric is the bottleneck, not the disks.
For OLTP, we don't care about Gigabytes; we care about Latency and Message Rate.
The Silicon Limit: A full rack has $\sim 1,500+$ CPU cores.
The RDMA Advantage: Because Exadata uses RDMA (Remote Direct Memory Access) for "Single Block Reads," the DB Server CPU doesn't participate in the I/O.
The Calculation: With sub-20$\mu$s RDMA reads, a single rack can realistically push over 30 Million IOPS and comfortably exceed 2–3 Million TPS, provided the "Redo Log" (the serial bottleneck) is placed on Persistent Memory (PMEM).
When performing heavy filtering (Predicate Offload), the throughput isn't measured by what hits the DB Server, but by what the Storage Cells process.
| Metric | Theoretical Max (Full Rack) | Bottleneck |
| Scan Throughput | 1.2 TB/s (Aggregated Cell Scan) | Storage CPU (Decompression) |
| Data to DB Node | 300 GB/s | RoCE Network Fabric |
| OLTP IOPS | 35,000,000+ | PCIe Bus Latency |
| Log Writes | 12-15 GB/s | PMEM/RDMA Persistence |
A theoretical model must account for the Fan-In Ratio. In a 1M TPS rack, you have 14 Storage Cells "feeding" 8 DB Nodes.
The Ratio: 1.75 : 1.
The Impact: This over-provisioning of storage ensures that even if two storage nodes suffer a "Rebalance Storm" or thermal throttling, the DB Nodes still receive enough data to maintain the 1M TPS target.
The theoretical maximum of an Exadata rack is defined by the Minimum of these three variables:
Fabric Limit: Total RoCE Bandwidth ($\approx 300\text{ GB/s}$).
Compute Limit: Total Cores $\times$ Context Switch Efficiency.
Redo Limit: The sequential write speed of the PMEM/Log Buffer.
A single Exadata X11M rack is theoretically capable of ~5 Million OLTP TPS or ~1 Terabyte/sec of Analytics Scan. However, in the real world, you hit the "Silicon-Level Concurrency Limits" (Cache Bouncing and Latch Contention) long before you hit the PCIe bus limits.
The true "Maximum" is achieved when you balance your workload so that the Storage Cells are 80% busy with Smart Scans while the DB Nodes are 80% busy with OLTP, utilizing the "Co-Designed" nature of the rack to its absolute limit.