3 ms·
Did you read the PDFs that are cited in this thread? Paul Turner's switch-to scheme switches between real threads orders of magnitude faster than the kernel sc
by romed 8y ago
Did you read the PDFs that are cited in this thread? Paul Turner's switch-to scheme switches between real threads orders of magnitude faster than the kernel scheduler. So it's not true that "you're not really going to beat the kernel by enough to matter". It's actually much faster.
- taeric 8y agoEven beating it by an order of magnitude doesn't automatically guarantee enough to matter, does it? As an analogy, I could speed up my ability to get shoes on by an order of magnitude, and have it not really matter for my commute. Not saying that is exactly the same here. Just extending the question on the impact if this.
- romed 8y agoThat's a fair point. Such schemes are for people whose profiles clearly implicate the kernel scheduler.
- taeric 8y agoDo you have a good link outlining who that likely is?
- romed 8y agoNo, I don’t, but in my experience at Google the kernel scheduler is a big problem for thread-per-request servers pushed to their limits. Async completion-passing servers are more efficient at their limits but more difficult to write, maintain, and debug. Userspace thread scheduling expands the performance envelope of thread-per-request architecture.
- taeric 8y agoThis fits my expectation. I still tell most folks that starting at thread per request is a sensible starting place. The difficulty in writing continuation style can cause most efforts to stall out. Of course, that last assertion needs data. And i could be wrong. :(