Fwd: [Haskell-cafe] For STM to practically serve large in-mem datasets with cyclic structures WAS: STM friendly TreeMap (or similar with range scan api) ? WAS: Best ways to achieve throughput, for large M:N ratio of STM threads, with hot TVar updates?

Dear GHC Devs, I realize I should seek your helps regarding my current situation. TL;DR the most serious suffering is: After the heap loaded with many TVars with cyclic structures, GC will dominate my CPU utilization with little business progressing. Nonmoving GC `-xn` with 8.10.1 helps, but the ceiling is not high enough for my case, obvious performance degrade starts at about ~350MB RSS, and falls unusable as RSS approaching 1GB. While I expect a GHC compiled process to serve 20~30GB in-mem data practically okay. https://mail.haskell.org/pipermail/haskell-cafe/2020-July/132556.html https://mail.haskell.org/pipermail/haskell-cafe/2020-July/132556.html contains most interesting conversations. I found https://tech.channable.com/posts/2020-04-07-lessons-in-managing-haskell-memo... https://tech.channable.com/posts/2020-04-07-lessons-in-managing-haskell-memo... much relevant, they managed to keep 100GB per instance, but switching to immutable data structures to reside within compact regions is not immediately feasible for my case, as the schema has to be re-designed, though I have my colleges started evaluating that option. Best regards, Compl

Begin forwarded message:

From: YueCompl via Haskell-Cafe Subject: [Haskell-cafe] For STM to practically serve large in-mem datasets with cyclic structures WAS: STM friendly TreeMap (or similar with range scan api) ? WAS: Best ways to achieve throughput, for large M:N ratio of STM threads, with hot TVar updates? Date: 2020-07-30 at 21:28:31 GMT+8 To: Haskell Cafe Reply-To: YueCompl

For the record, overhead of STM over IO (or other means where manual composition of transactions needed) based concurrency control, is a price I'm willing to pay in my use case, as it's not machine-performance critical in distributing input data + parameters to a cluster of worker nodes, and collecting their results into permanent storage or a data pipeline. But to keep professionally crafting well synced, race-free scheduling code is barely affordable by my org, as shape of datasets, relationship between them and algorithms processing them are varying at fast paces, we have difficulty, or lack the willingness, to hire some workforce specifically to keep each new data pipeline race free, it has to be, but better at cost of machine-hours instead of human head counts.

While easily compositing stm code, wrapped in scriptable procedures, will enable our analysts to author the scheduling scripts without too much concerns. Then our programmers can focus on performance critical parts of the data processing code, like optimization of tight-loops.

Only if not in the tight loops, I think it's acceptable by us, that up to 2~3 order of magnitude slower for an stm solution compared to its best rivals, as long as it's scalable. For a (maybe cheating) example, if fully optimized code can return result in 10 ms after an analyst clicked a button, we don't bother if unoptimized stm script needs 10 second, so long as the result is correct.

In a philosophic thinking, I heard that AT&T had UNIX specifically designed for their Control panel, while their Data panel runs separate software (and on separate hardware obviously), while modern systems have powerful CPUs tempting us to squeeze more performance out of it, and SIMD instructions make it even more tempting, I think we'd better resist it when programming something belong to the Control panel per se, but do it in programming something belong to the Data panel. And appears Data panel programs are being shifted to GPUs nowadays, which feels right.

Regards, Compl

...

On 2020-07-30, at 20:10, YueCompl via Haskell-Cafe mailto:haskell-cafe@haskell.org> wrote:

Hi Peter,

Great to hear from you!

For the record tskiplist (and stm-containers together) did improve my situation a great lot with respect to scalability at concurrency/parallelism!

I'm still far from the stage to squeeze last drops of performance, currently I'm just making sure performance wise concerns be reasonable during my PoC in correctness and ergonomics of my HPC architecture (an in-memory graph + out-of-core (mmap) array DBMS powered computation cluster, with shared storage), and after parallelism appears acceptable, I seemingly suffer from serious GC issue at up scaling on process working memory size. I'm suspecting it's because of the added more TVars and/or aggressive circular structures of them in my case, and can not find a way to overcome it by far.

Thanks for your detailed information!

Best regards, Compl

...

On 2020-07-30, at 19:19, Peter Robinson mailto:pwr@lowerbound.io> wrote:

Hi Compl,

...

+ This package provides a proof-of-concept implementation of a skip list in STM

This has to mean something but I can't figure out yet.

Dear Peter Robinson, I hope you can see this message and get in the loop of discussion.

The reason for adding this sentence was that tskiplist hasn't been optimized for production use. Later on, I wrote an implementation of a concurrent skip list with atomic operations that performs significantly better, but it's operations work in the IO monad.

I'm surprised to hear that you're getting poor performance even when using the stm-container package, which I believe was meant to be used in production. A while ago, I ran some benchmarks comparing concurrent dictionary data structures (such as stm-container) under various workloads. While STMContainers.Map wasn't as fast as the concurrent-hashtable package, the results indicate that the performance doesn't degrade too much under larger workloads.

You can find these benchmark results here (10^6 randomly generated insertion/deletion/lookup requests distributed among 32 threads): https://lowerbound.io/blog/bench2-32.html https://lowerbound.io/blog/bench2-32.html And some explanations about the benchmarks are here: https://lowerbound.io/blog/2019-10-24_concurrent_hash_table_performance.html https://lowerbound.io/blog/2019-10-24_concurrent_hash_table_performance.html

One issue that I came across when implementing the tskiplist package was this: If a thread wants to insert some item into the skip list, it needs to search for the entry point by performing readTVar operations starting at the list head. So, on average, a thread will read O(log n) TVars (assuming a skip list of n items) and, if any of these O(log n) TVars are modified by a simultaneously running thread, the STM runtime will observe a (false) conflict and rerun the transaction. It's not clear to me how to resolve this issue without access to something like unreadTVar (see [1]).

Best, Peter

[1] UnreadTVar: Extending Haskell Software Transactional Memory for Performance (2007) by Nehir Sonmez , Cristian Perfumo , Srdjan Stipic , Adrian Cristal , Osman S. Unsal , Mateo Valero.

_______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

_______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.

_______________________________________________ Haskell-Cafe mailing list To (un)subscribe, modify options or view archives go to: http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe Only members subscribed via the mailman list are allowed to post.