Insert performance

[fireya]

Hello!
I have noticed slow insert performance of insert operations in connector. To test this, i have made some benchmarks:
Create space:

box.schema.space.create('pivot')
box.space.pivot:create_index('primary', { type = 'tree', parts = {1, 'num'}, unique = true})
box.space.pivot:create_index('rtree', { type = 'rtree', parts = {2, 'array'}, unique = false})

Test on lua:

box.space.pivot:truncate()
function main_function()
  local t
  for i = 1,1000000,1 do
    t = box.tuple.new({i,{i,i},i})
    box.space.pivot:insert(t)
  end
end
start_time = os.clock()
main_function()
end_time = os.clock()
'insert done in ' .. end_time - start_time .. ' seconds'

Results: 4 seconds, or 250k per second

Test on go:
first, truncate space: box.space.pivot:truncate()
then test:

package main

import (
	"fmt"
	"log"
	"time"

	"github.com/tarantool/go-tarantool"
)

func main() {
	opts := tarantool.Opts{}
	conn, err := tarantool.Connect("10.40.10.143:3301", opts)

	// conn, err := tarantool.Connect("/path/to/tarantool.socket", opts)
	if err != nil {
		fmt.Println("Connection refused: %s", err.Error())
	}
	start := time.Now()
	f := make([]*tarantool.Future, 0)
	for i := 0; i < 1000000; i++ {
		fut := conn.InsertAsync("pivot", []interface{}{i, []int{i, i}, i})
		f = append(f, fut)
	}
	for _, element := range f {
		_, err := element.Get()
		if err != nil {
			fmt.Println("Insert failed: %s", err.Error())
		}
	}
	elapsed := time.Since(start)
	log.Printf("Insert took %s", elapsed)
}

Results: 6.74s or 148k per second

Now .net connector:
truncate: box.space.pivot:truncate()
Test:

using System;
using System.Diagnostics;
using System.Threading.Tasks;
using ProGaudi.Tarantool.Client;
using ProGaudi.Tarantool.Client.Model;

namespace Tarantool.Test
{
    class Program
    {
        static void Main(string[] args)
        {
            using (var box = Box.Connect("10.40.10.143:3301").Result)
            {
                var schema = box.GetSchema();
                var space = schema["pivot"];
                var sw = new Stopwatch();
                var lst = new Task[1000000];
                sw.Start();
                for (int i = 0; i < lst.Length; i++)
                {
                    lst[i] = space.Insert(new TarantoolTuple<int, int[], int>(i, new[] { i, i }, i));
                }

                Task.WaitAll(lst);
                sw.Stop();
                Console.WriteLine(sw.ElapsedMilliseconds);
                Console.ReadKey();
            }
        }
    }
}

Time: 92 sec or 11k per second

[aensidhe]

I have better results (33k per second), but worse than that go results. You can improve by at least 30-50% by batching queries into batches of 1000, right now you're benchmarking .net GC more than actual code of library.

That does not mean, that I will leave it as it is. I'll need to rewrite our or adopt new msgpack library, most problem lie there. Unfortunately, can't provide any ETA as hard promises.

This branch has some drafts of what I want to do with it.

[Totktonada]

Whether something changes after PR #147?

[Totktonada]

The related question on StackOverflow: https://stackoverflow.com/questions/69138615/what-is-the-best-way-to-upload-large-data-set-to-tarantool.

[adamlepkowski]

Hi, I'm the author of the post in stackoverflow that Totktonada mention above. In my case I have max 63 req per second. Below you can find memory statistics:

- items_size: 32203920
   items_used_ratio: 97.42%   !!
quota_size: 268435456   
quota_used_ratio: 18.75%   
arena_used_ratio: 82.1%   
items_used: 31373320   
quota_used: 50331648   
arena_size: 50331648   
arena_used: 41302024 ...

Those settings were defined:

 box.cfg{memtx_max_tuple_size=100*1048576}
 box.cfg{memtx_memory=3*268435456}

[karlovnv]

new Task[1000000];

It's not a good idea to run a such number of tasks. It's better to initialize several long running tasks with loops.
Something like this pseudocode:

var tasksList = new List<Task>();
for(int i = 0; i< 30; i++)
{
    taskList.Add(Task.Run( () => 
    {
         for(j = 0; j < 100000; j++)
         {
              space.Replace(new TarantoolTuple<int, int[], int>(j, new[] { j, j }, j)); // replace in order to avoid PK conflicts
         }
    }).ToArray();
}
//....
Task.WhenAll(tasksList).Wait();
//....