In Onur Mutlus Vorlesung wurde heute die TLB-Reichweite nicht erwähnt. Angesichts dessen verfehlt Intels TLB-Design den Zweck von Superpages, weil seine TLB-Reichweite in Stein gemeißelt ist. Deswegen ist diese Auslassung so wesentlich.
In Onur Mutlu's lecture today, TLB reach wasn't mentioned. Intel's TLB design defeats the entire point of superpages in light of that because its TLB reach is set in stone, and that's why the omission is so glaring.
map_anon_folio_pte_pf() et al that I see from post-7.0 code I have to merge my rolling pgcl port with appear to be helpers for dealing with multiple superpage sizes as part of the mTHP that this is the first I've heard anything of. That's moderately hopeful. With mTHP in hand, just pgcl would bring Linux' superpaging to parity with Telix', with various amounts of uncertainty as to how well either one is achieving promotions esp. in sub- allocation unit contexts.
For evaluation strategies, I see:
- 1. TLB-MPKI Talluri-Hill (ASPLOS '94), Navarro-Iyer-Druschel-Cox (OSDI '02), Barr-Cox-Rixner (ISCA '10)
- 2. cache pollution, maybe LLC-MPKI, originators unclear
- 3. LLC/TLB-MPKI vs. RAM scaling curves to determine „memory walls” per Wulf-McKee '95
- 4. Attribution / stack-distance modelling per Mattson '70
My own list of things I'd like to measure is:
- 1. Queueing: inter-arrival times and service times have more predictive power
- 2. Distribution fairness: the grand total never matters as much as the distribution, where fairness is a central concern, as well as priority inversion
- 3. Distribution utilisation: is effective use being made of sizes across the size spectrum?
- 4. Verification of small superpage size allocation guarantees
- 5. Fragmentation metrics
- 6. Verification of increased allocation success likelihoods for larger page sizes in the face of external fragmentation
- 7. Defragmentation overhead
- 8. Distribution to shared mappings: the highest impact can be had via sharing, so it's also a question of whether the contiguity has been distributed to the shared memory objects of greatest impact.
The MPKI notion is on the one hand a reciprocal of inter-arrival times and on the other isn't naturally comparable or extensible to service times. Wulf-McKee recast in queueing theoretical terms still doesn't cover the full range of issues I'm working on, though. Metrics for 4. distribution fairness, 7. defragmentation overhead and 8. distribution to shared mappings and maybe more need to be devised.
Nadia/Надя/नाडिया/娜迪亚/ نادية