With quickly enlargement of internet and cloud computing, warehouse-scale computing (WSC) workloads (search, email, video sharing, online maps, online shopping, etc.) have reached planetary scale and are driving the lion’s share of growth in computing demand. WSC workloads also differ from others in their requirements for on-demand scalability, elasticity and availability.
Many studies (e.g., Profiling a warehouse-scale computer) and books (e.g., The Datacenter as a Computer: Designing Warehouse-Scale Machines) have pointed out that WSC workloads have basically different traits than traditional benchmarks and require changes to modern computer architecture to achieve optimum efficiency. Google workloads have data and instruction footprints that go past the capacity of modern CPU caches, such that the CPU spends a significant portion of its time waiting for code and data. Simply increasing memory bandwidth would not solve the dispute, as many accesses are in the critical path for application request processing; it is just as necessary to reduce memory entry latency as it is to increase memory bandwidth.
Over the years, the computer architecture community has expressed the need for WSC workload traces to perform architecture research. Today, we are pleased to announce that we’ve revealed select Google workload traces. These traces will help systems designers better understand how WSC workloads perform as they interact with underlying components, and develop new solutions for entrance-end and data-entry bottlenecks.
We captured these workload traces using DynamoRIO on computer servers operating Google workloads — you can find more details at https://dynamorio.org/google_workload_traces.html. To protect user privacy, these traces only contain instruction and memory addresses.
We have found these traces useful for understanding WSC workloads and seeding internal research on processor entrance-ends, on-die interconnects, caches and memory subsystems, etc. — all areas that greatly impact WSC workloads. For example, we used these traces to develop AsmDB. Likewise, we hope these traces will enable the computer architecture community to develop new ideas that improve performance and efficiency of other WSC workloads.