Note this is not true random access in the manner it occurs in most programs. By having a contiguous array of indices to look at, that array can be prefetched as it goes, and speculative execution will take care of loading many upcoming indices of the target array in parallel.
A more interesting example might be if each slot in the target array has the next index to go to in addition to the value, then you will introduce a dependency chain preventing this from happening.
> A more interesting example might be if each slot in the target array has the next index to go to in addition to the value, then you will introduce a dependency chain preventing this from happening.
However, on some processors there's a data-dependent prefetcher that will notice the pointer-like value and start prefetching that address before the CPU requests it.
The data-dependent prefetcher is a cool feature, though you do have to be careful with side-channel issues, so some of them can disable it with the Data-Independent Timing bit or similar.
At this point I'm kinda expecting CPU vendors to stop putting as many Spectre mitigations in the main core, and just have a small crypto core with full-fat arithmetic, less hardware for memory access, less speculation, and careful side-channel hardening. You still have to block Meltdown and other large vulnerabilities on the main cores, but if someone wants to protect elliptic curves from weird attacks? Try to set the DIT bit, trap into the OS, and get sent to the hardened core.
Even if the prefetcher was capable of traversing pointers, it wouldn't help. The hypothetical benchmark wouldn't do anything other chasing pointers, and the prefetcher can't really do that any quicker. A traversing prefetcher is useful if the code actually does work for each traversed node, then the prefetcher (or the OoO machinery) could realistically run ahead.
Do you mean storing the next index/offset and having the pointer value calculated as late as possible by adding the starting address (and maybe multiplying the index by sizeof)? That would probably defeat/mislead Intel's prefetcher, as described at https://www.intel.com/content/www/us/en/developer/articles/t...
If the next index is stored in the target array and the indexes are random, you will likely get a cycle of length O(sqrt(n)), which can be cached.
You can avoid this with two arrays. One contains random query positions, and the target array is also filled with random values. The next index is then a function of the next query position and the previous value read from the target array.
You could sample from a different random distribution.
Eg start with every element referencing the next element (i.e. i+1 with wrap-around), and then use a random shuffle. That way, you preserve the full cycle.
You can do that, but it's inefficient with larger arrays. Iterating over 10 billion elements in a random order can take up to an hour. Which is probably more than what you are willing to wait for a single case in a benchmark. On the other hand, you will probably find a cycle in a uniformly random array of 10 billion elements within 0.1 seconds, which is not enough to mitigate the noise in the measurements. So you need a way of generating unpredictable query positions for at least a few seconds without wasting too much time setting it up.
You seem to misunderstand @andersa's point which I think is well expressed - it doesn't matter if the indices are randomized if the CPU can pre-fetch what they will be. The power of CPU speculative execution to hide latency can be quite surprising the first time you see it.
This is a very small Nim program to demonstrate for "show me the code" and "it must just not be 'random enough'!" skeptics: https://github.com/c-blake/bu/blob/main/memlat.nim It uses the exact dependency idea @andersa mentions of a random cycle of `x[i] = i` that others else-sub-thread say some CPUs these days are smart enough to "see through". On Intel CPUs I have, the dependency makes things 12x slower at the gigabyte scale (DIMMs).
EDIT: This same effect makes many a naive hash table microbenchmark { e.g., `for key in keys: lookup(key)` } unrepresentative of performance in real programs where each `key` is often not speculatively pre-computable.
In the end it depends exactly what you want to measure. Of course a load-load dependency will make everything as slow as the latency of the cache level you are accessing as that becomes the bottleneck.
Traversing a contiguous list of pointers in L1 is also slower than accessing those pointers by generating their address sequentially, so adding a load-load dependency is not a good way to benchmarking random access vs sequential access (it is a good way to benchmark vector traversal vs list traversal of course).
At the end of the day you have to accept that like caching and prefetching speedup sequential access, OoO execution[1] will speedup (to a lesser extent) random access. Instead of memory latency, in this case the bottleneck would be the OoO queue depth, or more likely the maximum number of outstanding L1/L2/L3 (and potentially TLB) misses. As long as the maximum number of outstanding misses is lower than the memory latency for that cache level, then, in first approximation, the cpu can effectively hide the sequential vs random access cost for independent accesses.
Benchmarking is hard. Making sure that that a microbenchmark represents your load effectively, doubly so.
[1] Even many in-order CPUs have some run-ahead capabilities for memory.
All true. No real disagreement and I'm often saying "it all depends.." myself. :-) In this case, there is also some vagueness around "random" (predictable to what subsystem when).
I still suspect @kortilla is one of today's lucky 10,000 (https://xkcd.com/1053/) or just read/replied too quickly. :-)
There is a lot written that indicates that the complexity of modern CPUs is ill-disseminated. But there is also wonderful stuff like https://gamozolabs.github.io/metrology/2019/08/19/sushi_roll... { To add a couple links to underwrite my reply in full agreeance. :-) }
Intel added PREFETCHW to their Broadwell processors launched in 2014, years after AMD dropped all 3DNow! instructions except the prefetch instructions. That timeline strongly suggests that the instructions aren't no-ops and likely are used by some popular software.
Hardware definitely supports this but it might need compiler support, as in adding instructions to do prefetching. Which might be done automatically or requires a pragma or calling a builtin. So it can be implemented in any case.
What surprises me is the 24 GB of DDR4 DRAM on a dual channel memory controller? AFAIK there are only 8 GB or 16 GB modules, no 12 GB modules. At least I can only find 12 GB DDR5 modules listed, but not DDR4.
This means: The system likely uses 3x 8 GB modules. As a result, one channel has two modules with 16 GB total, while the other channel has only a single 8 GB module.
Not sure how big this impact is with the given memory access patterns and assuming [mostly] exclusive single-threaded access. It's just something I noted, and could be a source of unexpected artifacts.
The RandomAccess (or GUPS) benchmark (see: https://ieeexplore.ieee.org/document/4100365) was looking at measuring machines on this kind of workload. In high performance computing this was important for graph calculations and was one of the things the Cray (formerly Tera) MTA machine was particularly good at. I suppose this benchmark wouldn’t be very widely known outside HPC circles.
I worked on the MTA architectures for years among several other HPC systems but I don’t remember this particular benchmark. I suspect it was replaced by the Graph500 benchmark. Graph500 measures something similar and was introduced only a few years after GUPS.
The HPCS benchmarks predated Graph500. They were talked about at SC for a few years in the early 2000s but mostly faded into the background. It’s hard to dig up the numbers for the MTA on RandomAccess, but the Eldorado paper from ‘05 by Feo and friends (https://dl.acm.org/doi/10.1145/1062261.1062268) mentions it and you can see the MTA beating the other popular architectures of the time in one of the tables.
Feo was a major MTA stan and proponent, even years later. Honestly, it is probably my favorite computing architecture of all time despite the weaknesses of the implementation. It was extraordinarily efficient in some contexts. Few people could design properly optimized code for them though, which was an additional problem.
There were proofs of concept by 2010 that the latency-hiding mechanics could be implemented on CPUs in software, which while not as efficient had the advantage of cost and performance, which was a death knell for the MTA. A few attempts to revive that style of architecture have come and gone. It is very difficult to compete with the economics of mass-scale commodity silicon.
I hold out hope that a modern barrel processor will become available at some point but I’m not sanguine about it.
Random access is catastrophically slower because of the successive cache misses when the prefetcher fails to guess what you're doing.
One hint in the same article that random access is not cheap, in contrast with the conclusion, was noticing that the shuffle was unacceptably slow on large data sets.
Still, good to see peformance measurements, especially where the curves look roughly like you'd hope them to.
longtime back ran a model scoring job where each row did feature lookups across diff memory regions. access pattern changed every run since it depended on prev feature values. avg latency stayed fine but total run time kept jumping sometimes 6s, sometimes 8s on same input. perf counters looked flat but thread scheduling kept shifting. reordered the lookups to cut down mem jumps. after that threads aligned better, scheduler ran smoother, run time got stable across batches. gain wasn’t from faster mem, imo i observed it's not the memory is slower but rather how less predictable it's
I did another type of experiment which evaluates benefits of branch prediction on AMD 9950X on contiguous array with 1,000,000 elements. Calculated sum adding element if it is bigger than 125 (50% of 256). Difference between random and sorted was 10 times. I guess branch prediction plays a huge role as well.
Presumably if you'd split the elements into 16 shares (one for each CPU), summed with 16 threads, and then summed the lot at the end, then random would be faster than sorted?
I don’t think random should be faster than contiguous access, if you parallelize both of them.
Although, it looks like that chip has a 1MB L2 cache for each core. If these are 4 Bytes ints, then I guess they won’t all fit in one core’s L2, but maybe they can all start out in their respective cores’ L2 if it is parallelized (well, depends on how you set it up).
Maybe it will be closer. Contiguous should still win.
What are your qualms with time per element? I liked it as a metric because it kept the total deviation of results to less than 32 across the entire result set.
Using something like the overall run length would have such large variations making only the shape of the graph particularly useful (to me) less so much the values themselves.
If I was showing a chart like this to "leadership" I'd show with the overall run length. As I'd care more about them realizing the "real world" impact rather than the per unit impact. But this is written for engineers, so I'd expect it to also be focused on per unit impacts for a blog like this.
However, having said all that, I'd love to hear what your reservations are using it as a metric.
I also just don’t know what to do with “1 ns per element”. The scale of 1 to 4 ns per element is remarkably imprecise. Discussing 1 to 250 million to 1 billion elements per second feels like a much wider range. Even if it’s mathematically identical.
Your graphs have a few odd spikes that weren’t deeply discussed. If it’s under 2ns per element who cares!
The logarithmic scale also made it really hard to interpret. Should have drawn clearer lines at L1/L2/L3/ram limits.
On skim I don’t think there’s anything wrong. But as presented it’s a little hard for me as an engineer to extract lessons or use this information for good (or evil).
There shouldn’t be a Linux vs Mac issue. Ignoring mmap this should be HW.
I dunno. Those are all just surface level reactions.
Lots of things are expected when you deeply understand a complex system and think about it. But, like, not everyone knows the system that deeply nor have they thought about it!
Note this is not true random access in the manner it occurs in most programs. By having a contiguous array of indices to look at, that array can be prefetched as it goes, and speculative execution will take care of loading many upcoming indices of the target array in parallel.
A more interesting example might be if each slot in the target array has the next index to go to in addition to the value, then you will introduce a dependency chain preventing this from happening.
> A more interesting example might be if each slot in the target array has the next index to go to in addition to the value, then you will introduce a dependency chain preventing this from happening.
However, on some processors there's a data-dependent prefetcher that will notice the pointer-like value and start prefetching that address before the CPU requests it.
The data-dependent prefetcher is a cool feature, though you do have to be careful with side-channel issues, so some of them can disable it with the Data-Independent Timing bit or similar.
At this point I'm kinda expecting CPU vendors to stop putting as many Spectre mitigations in the main core, and just have a small crypto core with full-fat arithmetic, less hardware for memory access, less speculation, and careful side-channel hardening. You still have to block Meltdown and other large vulnerabilities on the main cores, but if someone wants to protect elliptic curves from weird attacks? Try to set the DIT bit, trap into the OS, and get sent to the hardened core.
Even if the prefetcher was capable of traversing pointers, it wouldn't help. The hypothetical benchmark wouldn't do anything other chasing pointers, and the prefetcher can't really do that any quicker. A traversing prefetcher is useful if the code actually does work for each traversed node, then the prefetcher (or the OoO machinery) could realistically run ahead.
Could probably overcome that by using integers, but converting them to a pointer after accessing (like '0'+1 is '1').
Do you mean storing the next index/offset and having the pointer value calculated as late as possible by adding the starting address (and maybe multiplying the index by sizeof)? That would probably defeat/mislead Intel's prefetcher, as described at https://www.intel.com/content/www/us/en/developer/articles/t...
Fun fact, that's part of why parsing protobuf is so slow.
If the next index is stored in the target array and the indexes are random, you will likely get a cycle of length O(sqrt(n)), which can be cached.
You can avoid this with two arrays. One contains random query positions, and the target array is also filled with random values. The next index is then a function of the next query position and the previous value read from the target array.
You could sample from a different random distribution.
Eg start with every element referencing the next element (i.e. i+1 with wrap-around), and then use a random shuffle. That way, you preserve the full cycle.
You can do that, but it's inefficient with larger arrays. Iterating over 10 billion elements in a random order can take up to an hour. Which is probably more than what you are willing to wait for a single case in a benchmark. On the other hand, you will probably find a cycle in a uniformly random array of 10 billion elements within 0.1 seconds, which is not enough to mitigate the noise in the measurements. So you need a way of generating unpredictable query positions for at least a few seconds without wasting too much time setting it up.
This is why array random access and linked-list random access have wildly different performance characteristics.
Another thing I noticed is that the spike on the left hand side of his graphs is the overhead of file access.
Without this overhead, small array random access should have a lot better per-element cost.
The article very clearly compares using randomized indexes and sequential. It’s kinda the point of the article.
You seem to misunderstand @andersa's point which I think is well expressed - it doesn't matter if the indices are randomized if the CPU can pre-fetch what they will be. The power of CPU speculative execution to hide latency can be quite surprising the first time you see it.
This is a very small Nim program to demonstrate for "show me the code" and "it must just not be 'random enough'!" skeptics: https://github.com/c-blake/bu/blob/main/memlat.nim It uses the exact dependency idea @andersa mentions of a random cycle of `x[i] = i` that others else-sub-thread say some CPUs these days are smart enough to "see through". On Intel CPUs I have, the dependency makes things 12x slower at the gigabyte scale (DIMMs).
EDIT: This same effect makes many a naive hash table microbenchmark { e.g., `for key in keys: lookup(key)` } unrepresentative of performance in real programs where each `key` is often not speculatively pre-computable.
In the end it depends exactly what you want to measure. Of course a load-load dependency will make everything as slow as the latency of the cache level you are accessing as that becomes the bottleneck.
Traversing a contiguous list of pointers in L1 is also slower than accessing those pointers by generating their address sequentially, so adding a load-load dependency is not a good way to benchmarking random access vs sequential access (it is a good way to benchmark vector traversal vs list traversal of course).
At the end of the day you have to accept that like caching and prefetching speedup sequential access, OoO execution[1] will speedup (to a lesser extent) random access. Instead of memory latency, in this case the bottleneck would be the OoO queue depth, or more likely the maximum number of outstanding L1/L2/L3 (and potentially TLB) misses. As long as the maximum number of outstanding misses is lower than the memory latency for that cache level, then, in first approximation, the cpu can effectively hide the sequential vs random access cost for independent accesses.
Benchmarking is hard. Making sure that that a microbenchmark represents your load effectively, doubly so.
[1] Even many in-order CPUs have some run-ahead capabilities for memory.
All true. No real disagreement and I'm often saying "it all depends.." myself. :-) In this case, there is also some vagueness around "random" (predictable to what subsystem when).
I still suspect @kortilla is one of today's lucky 10,000 (https://xkcd.com/1053/) or just read/replied too quickly. :-)
There is a lot written that indicates that the complexity of modern CPUs is ill-disseminated. But there is also wonderful stuff like https://gamozolabs.github.io/metrology/2019/08/19/sushi_roll... { To add a couple links to underwrite my reply in full agreeance. :-) }
> By having a contiguous array of indices to look at, that array can be prefetched as it goes
Does x86 64 actually do this data dependent single deref prefetech? Because in that case I have a some design assumptions I have to reevaluate.
The compiler probably does [0].
[0] https://gcc.gnu.org/projects/prefetch.html
That list doesn't include any current mainline processors. It's all Itanium, 3DNow!, and MIPS.
Intel added PREFETCHW to their Broadwell processors launched in 2014, years after AMD dropped all 3DNow! instructions except the prefetch instructions. That timeline strongly suggests that the instructions aren't no-ops and likely are used by some popular software.
Hardware definitely supports this but it might need compiler support, as in adding instructions to do prefetching. Which might be done automatically or requires a pragma or calling a builtin. So it can be implemented in any case.
What surprises me is the 24 GB of DDR4 DRAM on a dual channel memory controller? AFAIK there are only 8 GB or 16 GB modules, no 12 GB modules. At least I can only find 12 GB DDR5 modules listed, but not DDR4.
This means: The system likely uses 3x 8 GB modules. As a result, one channel has two modules with 16 GB total, while the other channel has only a single 8 GB module.
Not sure how big this impact is with the given memory access patterns and assuming [mostly] exclusive single-threaded access. It's just something I noted, and could be a source of unexpected artifacts.
The RandomAccess (or GUPS) benchmark (see: https://ieeexplore.ieee.org/document/4100365) was looking at measuring machines on this kind of workload. In high performance computing this was important for graph calculations and was one of the things the Cray (formerly Tera) MTA machine was particularly good at. I suppose this benchmark wouldn’t be very widely known outside HPC circles.
I worked on the MTA architectures for years among several other HPC systems but I don’t remember this particular benchmark. I suspect it was replaced by the Graph500 benchmark. Graph500 measures something similar and was introduced only a few years after GUPS.
The HPCS benchmarks predated Graph500. They were talked about at SC for a few years in the early 2000s but mostly faded into the background. It’s hard to dig up the numbers for the MTA on RandomAccess, but the Eldorado paper from ‘05 by Feo and friends (https://dl.acm.org/doi/10.1145/1062261.1062268) mentions it and you can see the MTA beating the other popular architectures of the time in one of the tables.
Feo was a major MTA stan and proponent, even years later. Honestly, it is probably my favorite computing architecture of all time despite the weaknesses of the implementation. It was extraordinarily efficient in some contexts. Few people could design properly optimized code for them though, which was an additional problem.
There were proofs of concept by 2010 that the latency-hiding mechanics could be implemented on CPUs in software, which while not as efficient had the advantage of cost and performance, which was a death knell for the MTA. A few attempts to revive that style of architecture have come and gone. It is very difficult to compete with the economics of mass-scale commodity silicon.
I hold out hope that a modern barrel processor will become available at some point but I’m not sanguine about it.
Random access is catastrophically slower because of the successive cache misses when the prefetcher fails to guess what you're doing.
One hint in the same article that random access is not cheap, in contrast with the conclusion, was noticing that the shuffle was unacceptably slow on large data sets.
Still, good to see peformance measurements, especially where the curves look roughly like you'd hope them to.
Would a better benchmark not just use some kind of pseudo randomly generated sequence to avoid having two arrays?
If, of course, you have the CPU and its caches all to yourself.
longtime back ran a model scoring job where each row did feature lookups across diff memory regions. access pattern changed every run since it depended on prev feature values. avg latency stayed fine but total run time kept jumping sometimes 6s, sometimes 8s on same input. perf counters looked flat but thread scheduling kept shifting. reordered the lookups to cut down mem jumps. after that threads aligned better, scheduler ran smoother, run time got stable across batches. gain wasn’t from faster mem, imo i observed it's not the memory is slower but rather how less predictable it's
Love this analysis! Was expecting random to be much slower. 4x is not bad at all
There has to be some power hit for all those extra cache fills. No idea if it would be measurable.
is this not just a memory test for the burst capacitiy and access strategy of the dram controller?
I did another type of experiment which evaluates benefits of branch prediction on AMD 9950X on contiguous array with 1,000,000 elements. Calculated sum adding element if it is bigger than 125 (50% of 256). Difference between random and sorted was 10 times. I guess branch prediction plays a huge role as well.
Thanks for sharing that.
Presumably if you'd split the elements into 16 shares (one for each CPU), summed with 16 threads, and then summed the lot at the end, then random would be faster than sorted?
I don’t think random should be faster than contiguous access, if you parallelize both of them.
Although, it looks like that chip has a 1MB L2 cache for each core. If these are 4 Bytes ints, then I guess they won’t all fit in one core’s L2, but maybe they can all start out in their respective cores’ L2 if it is parallelized (well, depends on how you set it up).
Maybe it will be closer. Contiguous should still win.
Hm, no discussion of cache line size, page size, or the limits of cache associativity?
Here’s an older blog post of mine on roughly the same topic:
https://www.forrestthewoods.com/blog/memory-bandwidth-napkin...
I’m not sure I agree with the data presentation format. “time per element” doesn’t seem like the right metric.
What are your qualms with time per element? I liked it as a metric because it kept the total deviation of results to less than 32 across the entire result set.
Using something like the overall run length would have such large variations making only the shape of the graph particularly useful (to me) less so much the values themselves.
If I was showing a chart like this to "leadership" I'd show with the overall run length. As I'd care more about them realizing the "real world" impact rather than the per unit impact. But this is written for engineers, so I'd expect it to also be focused on per unit impacts for a blog like this.
However, having said all that, I'd love to hear what your reservations are using it as a metric.
It’s not wrong per se. I’m just very wary of nano-scale benchmarks. And I think in general you should advertise “velocity” not “time per”.
Perhaps it’s a long time inspiration from this post: https://randomascii.wordpress.com/2018/02/04/what-we-talk-ab...
I also just don’t know what to do with “1 ns per element”. The scale of 1 to 4 ns per element is remarkably imprecise. Discussing 1 to 250 million to 1 billion elements per second feels like a much wider range. Even if it’s mathematically identical.
Your graphs have a few odd spikes that weren’t deeply discussed. If it’s under 2ns per element who cares!
The logarithmic scale also made it really hard to interpret. Should have drawn clearer lines at L1/L2/L3/ram limits.
On skim I don’t think there’s anything wrong. But as presented it’s a little hard for me as an engineer to extract lessons or use this information for good (or evil).
There shouldn’t be a Linux vs Mac issue. Ignoring mmap this should be HW.
I dunno. Those are all just surface level reactions.
From your blog post:
> Random access from the cache is remarkably quick. It's comparable to sequential RAM performance
That's actually expected once you think about it, it's a natural consequence of prefetching.
If that wasn't the case the machine would have to prefetch to register file. I don't know of any CPU that does that.
Heh. That line often gets called out.
Lots of things are expected when you deeply understand a complex system and think about it. But, like, not everyone knows the system that deeply nor have they thought about it!
Whats most misleading is the data for the smaller sizes (1k)