On (25/01/28 17:21), Yosry Ahmed wrote: > On Tue, Jan 28, 2025 at 06:38:35PM +0900, Sergey Senozhatsky wrote: > > On (25/01/28 14:29), Sergey Senozhatsky wrote: > > > Maybe copy-in is just an okay thing to do. Let me try to measure. > > Yeah seems like the optimization is effective, at least on that > workload The workload is just lots of browser tabs (zram is configured as a swap device), so it's quite representative. > unless the memcpy() is cheap and avoiding it is not buying as > much > > ... > > (do you know if that's the case?). We run on arm64 and x86_64 on a variety of models from low-cost ones to high-cost ones. I probably wouldn't expect memcpy() of 6GB of random sized objects (loop unrolling is possible only partially) to be cheap in general.
