Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I'll give you a recent example: I've been offloading sparse experts to the CPU on a dual socket Zen system. These processors have multiple CPU dies, which connect through shared CCDs, which contain the L3 cache and the bus lines into DRAM and PCIe. (Additionally in the dual socket architecture there is a fabric which links the sockets.) And so by characterizing the physical architecture of the CPU and placing jobs accordingly, I can extract the maximum amount of performance from the system I have on hand.

Dell and some of the other hyperscalers have characterized these systems over the years in terms of maximum throughput via both the cache and the DRAM bus (and I have met or exceeded their numbers). But, nobody has actually characterized (and/or documented) the way that L3 cache eviction and retention works. Using a combination of fuzzing, patent research, and performance analysis, I have been able to characterize the operation of the system and tailor my stack accordingly.

Is this churn? Nah, this hardware is many years obsolete for mainstream deployment. Will this benefit me in my profession? Time will tell. Regardless, I had fun because I enjoy learning.

 help



Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: