With fast enlargement of web and cloud computing, warehouse-scale computing (WSC) workloads (search, e-mail, video sharing, on-line maps, on-line purchasing, and so forth.) have reached planetary scale and are driving the lion’s share of progress in computing demand. WSC workloads additionally differ from others of their necessities for on-demand scalability, elasticity and availability.
Many research (e.g., Profiling a warehouse-scale laptop) and books (e.g., The Datacenter as a Laptop: Designing Warehouse-Scale Machines) have identified that WSC workloads have essentially completely different traits than conventional benchmarks and require modifications to trendy laptop structure to realize optimum effectivity. Google workloads have information and instruction footprints that transcend the capability of recent CPU caches, such that the CPU spends a good portion of its time ready for code and information. Merely growing reminiscence bandwidth wouldn’t remedy the issue, as many accesses are within the important path for software request processing; it’s simply as essential to cut back reminiscence entry latency as it’s to extend reminiscence bandwidth.
Through the years, the pc structure neighborhood has expressed the necessity for WSC workload traces to carry out structure analysis. Right now, we’re happy to announce that we’ve printed choose Google workload traces. These traces will assist methods designers higher perceive how WSC workloads carry out as they work together with underlying parts, and develop new options for front-end and data-access bottlenecks.
We captured these workload traces utilizing DynamoRIO on laptop servers operating Google workloads — you will discover extra particulars at https://dynamorio.org/google_workload_traces.html. To guard consumer privateness, these traces solely comprise instruction and reminiscence addresses.
Now we have discovered these traces helpful for understanding WSC workloads and seeding inside analysis on processor front-ends, on-die interconnects, caches and reminiscence subsystems, and so forth. — all areas that drastically influence WSC workloads. For instance, we used these traces to develop AsmDB. Likewise, we hope these traces will allow the pc structure neighborhood to develop new concepts that enhance efficiency and effectivity of different WSC workloads.