This is very much worth watching. It is a tour de force.
Laurie does an amazing job of reimagining Google's strange job optimisation technique (for jobs running on hard disk storage) that uses 2 CPUs to do the same job. The technique simply takes the result of the machine that finishes it first, discarding the slower job's results... It seems expensive in resources, but it works and allows high priority tasks to run optimally.
Laurie re-imagines this process but for RAM!! In doing this she needs to deal with Cores, RAM channels and other relatively undocumented CPU memory management features.
She was even able to work out various undocumented CPU/RAM settings by using her tool to find where timing differences exposed various CPU settings.
>> It replicates data across multiple, independent DRAM channels with uncorrelated refresh schedules
This is the sort of thing which was done before in a world where there was NUMA, but that is easy. Just task-set and mbind your way around it to keep your copies in both places.
The crazy part of what she's done is how to determine that the two copies don't get get hit by refresh cycles at the same time.
Particularly by experimenting on something proprietary like Graviton.
Love the format, and super cool to see a benchmark that so clearly shows DRAM refresh stalls! Ran it on my 9950X3D machine with dual-channel DDR5 and saw clear spikes from 70ns to 330ns every 15us or so.
The hedging technique is quite a cool demo too, but I’m not sure it’s practical.
At a high level it’s a bit contradictory; trying to reduce the tail latency of cold reads by doubling the cache footprint makes every other read even colder.
I understand the premise is “data larger than cache” given the clflush, but even then you’re spending 2x the memory bandwidth and cache pressure to shave ~250ns off spikes that only happen once every 15us. There’s just not a realistic scenario where that helps.
Especially HFT is significantly more complex than a huge lookup table in DRAM. In the time you spend doing a handful of 70ns DRAM reads, your competitor has done hundreds of reads from cache and a bunch of math. It’s just far better to work with what you can fit in cache. And to shrink what doesn’t as much as possible.
Halfway through this great video and I have two questions:
1) Can we take this library and turn it into a a generic driver or something that applies the technique to all software (kernel and userspace) running on the system? i.e. If I want to halve my effective memory in order to completely eliminate the tail latency problem, without having to rewrite legacy software to implement this invention.
2) What model miniature smoke machine is that? I instruct volunteer firefighters and occasionally do scale model demos to teach ventilation concepts. Some research years back led me to the "Tiny FX" fogger which works great, but it's expensive and this thing looks even more convenient.
Laurie does an amazing job of reimagining Google's strange job optimisation technique (for jobs running on hard disk storage) that uses 2 CPUs to do the same job. The technique simply takes the result of the machine that finishes it first, discarding the slower job's results... It seems expensive in resources, but it works and allows high priority tasks to run optimally.
Laurie re-imagines this process but for RAM!! In doing this she needs to deal with Cores, RAM channels and other relatively undocumented CPU memory management features.
She was even able to work out various undocumented CPU/RAM settings by using her tool to find where timing differences exposed various CPU settings.
She's turned "Tailslayer" into a lib now, available on Github, https://github.com/LaurieWired/tailslayer
You can see her having so much fun, doing cool victory dances as she works out ways of getting around each of the issues that she finds.
The experimentation, explanation and graphing of results is fantastic. Amazing stuff. Perhaps someone will use this somewhere?
As mentioned in the YT comments, the work done here is probably a Master's degrees worth of work, experimentation and documentation.
Go Laurie!
This is the sort of thing which was done before in a world where there was NUMA, but that is easy. Just task-set and mbind your way around it to keep your copies in both places.
The crazy part of what she's done is how to determine that the two copies don't get get hit by refresh cycles at the same time.
Particularly by experimenting on something proprietary like Graviton.
Tis just probabilities and unlikelihood of hitting a refresh cycle across that many memory channels all at once.
The hedging technique is quite a cool demo too, but I’m not sure it’s practical.
At a high level it’s a bit contradictory; trying to reduce the tail latency of cold reads by doubling the cache footprint makes every other read even colder.
I understand the premise is “data larger than cache” given the clflush, but even then you’re spending 2x the memory bandwidth and cache pressure to shave ~250ns off spikes that only happen once every 15us. There’s just not a realistic scenario where that helps.
Especially HFT is significantly more complex than a huge lookup table in DRAM. In the time you spend doing a handful of 70ns DRAM reads, your competitor has done hundreds of reads from cache and a bunch of math. It’s just far better to work with what you can fit in cache. And to shrink what doesn’t as much as possible.
1) Can we take this library and turn it into a a generic driver or something that applies the technique to all software (kernel and userspace) running on the system? i.e. If I want to halve my effective memory in order to completely eliminate the tail latency problem, without having to rewrite legacy software to implement this invention.
2) What model miniature smoke machine is that? I instruct volunteer firefighters and occasionally do scale model demos to teach ventilation concepts. Some research years back led me to the "Tiny FX" fogger which works great, but it's expensive and this thing looks even more convenient.