In the past I discovered we do a lot of reposts while reading chunks of payload from the cache. It has major bad influence on reading performance. I was hoping we would need to fix this as one of optimizations after the new cache is turned on, but seems like this is too much of a regression. Attached is a detailed log snippet from a cache2 enabled talos run. I'd like to try to at least slightly increase performance here with good gain/risk ratio.

Talos results shows enlarging the chunk is enough for now. Also, I now recall I was playing with removing/minimizing the number of posts before we have the prioritization in the IO thread and I also found low influence of it to my page load tests when the IO prioritization was in the tree. Lowering impact rate here.