Hi Aaron, On Tue, Sep 25, 2018 at 1:04 PM Aaron West <aa...@loadbalancer.org> wrote: > It seems that the Kernel developers decided to halve the default TCP > memory in the 4.x kernels
Your colleague emailed the list about this last Nov. It was the ONLY thing I could find on this matter anywhere and was helpful in pointing me the right way. The crazy thing is that I doubled those numbers to what they were in Jessie and we still had slow downloads. I think this was because memory pressure mode was still being reached. Something else must have changed because I never touched tcp_mem prior to this and have never seen this sort of thing happen before. > Simply decide if you need to increase it by looking out for > the error message: Normally you won't ever see an error message, at least not in my experience. That's what was so frustrating about this. Once the middle value (pressure) is reached, the kernel appears to start throttling connections somehow (I think it starts to reduce the max buffer size that can be allocated per connection). Nothing is ever reported in the logs about this. Only when you set the three values for tcp_mem the same will you see the error message (at least the pressure and high values). > Anyway, just thought I'd mention it for info and to say you are not alone ;) Thanks, I appreciate it! -- Brendon Colby Senior DevOps Engineer Newgrounds.com