Benchmarks | ECP-CANDLE Benchmarks | Machine Learning library

I am in search of performance benchmarks for querying parquet ADLS files with the standard dedicated sql pool using external tables with polybase vs. serverless sql pool and OPENROWSET views. From my base queries on a 1.5 billion record table, it does appears OPENROWSET in serverless sql pool is around 30% more performant given time for the same query, but what are the architecture that power that? Are there any readily available performance benchmarks?

I have problem with BenchmarkDotNet which I struggle to solve

Here's my project structure:

I am trying to render a chart but encounter a problem: The elements appear in browsers (Chrome, Firefox) but not in traditional image viewers (Eyes of GNOME, GIMP, Inkscape).

Code

At first, I thought that it was because image viewers are incapable of rendering fonts, until I came across an asciinema's thumbnail, which is displayed perfectly by Eyes of GNOME:

Question: Why does this happen and how to fix this?

We have a setup where we are running 6 PgBouncer processes and our performance benchmarks degrade linearly with time. The longer PgBouncer has been running, the longer the connections to Postgres exist results in slower response times for the benchmark. We have a multi-tenant schema separated database with 2000+ relations. We are configured for Transaction Mode pooling right now. Over time, we see the memory footprint of each Postgres process climb and climb and climb, and again, this results in poorer performance.

We have tried to be more aggressive in cleaning up idle connections with the following settings:

I have a slice of strings. What I need to accomplish is to remove one value from the slice, without knowing the index. I thought this would be the easiest way to do it:

I wrote a function to add up all the elements of a double[] array using SIMD (System.Numerics.Vector) and the performance is worse than the naïve method.

On my computer Vector.Count is 4 which means I could create an accumulator of 4 values and run through the array adding up the elements by groups.

For example a 10 element array, with a 4 element accumulator and 2 remaining elements I would get

Having an intent to study a sort algorithm (of my own), I decided to compare its performance with the classical quicksort and to my great surprise I've discovered that the time taken by my implementation of quicksort is far not proportional to N log(N). I thoroughly tried to find an error in my quicksort but unsuccessfully. It is a simple version of the sort algorithm working with arrays of Integer of different sizes, filled with random numbers, and I have no idea, where the error can sneak in. I have even counted all the comparisons and swaps executed by my code, and their number was rather fairly proportional to N log(N). I am completely confused and can't understand the reality I observe. Here are the benchkmark results for sorting arrays of 1,000, 2,000, 4,000, 8,000 and 16,000 random values (measured with JMH):

I am reviewing some code and trying to come up with a technical reason why you should or should not use Task.WhenAll(Tasks[]) for essentially making Http calls in parallel. The Http calls call a different microservice and I guess one of the calls may or may not take some time to execute... (I guess I am not really interested in that). I'm using BenchmarkDotNet to give me an idea of there is any more memory consumed, or if execution time is wildly different. Here is an over-simplified example of the Benchmarks:

I have huge dynamically created tables/matrices in MATLAB of varying first dimension, whose rows represent (sorted) combinations of integers in the range 1-50 of order 6.

I would like to assign to each combination a unique value (hash, ranking), so that I can check if the same combinations appear in different tables. Different combinations are not allowed to have same value assigned, i.e. no collisions. I have to make a lot of such comparisons between a lot of such tables. So, for performance reasons, I would like to accomplish this by vectorization of uint32 operations to make it suitable for GPU acceleration in MATLAB.

Things I have thought of so far:

1. Lexicographic ranking: no idea how to vectorize the standard fast recursive algorithms well, and the only option seems to be to parfor it through the rows, which is slower than other options. IIRC, the direct explicit formula, though vectorizable, requires computation of binomials, which in turn requires log Gamma function in order to avoid huge factorials + double type to avoid collision if I am not mistaken, i.e. is slower because it's 'very numerical'.
2. Cantor pairing function: one can successively apply Cantor's pairing, which is nice because it's a polynomial expression, but it produces huge numbers well beyond uint32 and is definitely slower than other options.
3. Base 51 (no pun intended) integers: sends a combination/row vector (x_1,...,x_6) to x_1 + x_2 * 51 + ... + x_6 * 51^5. This is the fastest I currently have. It's easily vectorizable, but unfortunately still requires uint64 or double for rank-6 combinations of 50 elements, which is slower than uint32 or single type operations would be.

So, I guess, I am looking for a 'clever' injective function on these combinations that computes within the uint32 range and is also well vectorizable (in MATLAB).

Any help would be much appreciated!

EDIT: Here is a routine that benchmarks both ranking and searching in uint32, single, and double. I have used MATLAB's gputimeit to produce accurate results.