This gist shows benefits of table buckets preallocation

lj_table_new_nhash_bench.lua

jit.on()

local stopwatch = require('common.stopwatch')

local table_new = table.new

local attempts = 20
local keys = 100
local results = {}
local cached_keys = require('fun').range(keys):map(tostring):totable()

local function test1(iterations)
    local t = table_new(0, keys)
    for _ = 1, iterations do
        for j = 1, keys do
            t[cached_keys[j]] = j
        end
    end
    return t
end

local function test2(iterations)
    local t = nil
    for _ = 1, iterations do
        t = {}
        for j = 1, keys do
            t[cached_keys[j]] = j
        end
    end
    return t
end

local function test3(iterations)
    local t = nil
    for _ = 1, iterations do
        t = table_new(0, 0)
        for j = 1, keys do
            t[cached_keys[j]] = j
        end
    end
    return t
end

local function bench(iterations)
    results = table_new(attempts, 0)
    for _ = 1, attempts do
        table.insert(results, stopwatch.measure(function() test1(iterations) end, {format = 'ms'}))
    end
    table.sort(results)
    local median = results[math.floor(attempts / 2)]
    local operation_time = (median / iterations) * 1e6
    print(string.format("test1 - t = table.new(0, %d) median: %.2fms, %d insertions - %.2fns/op ", keys, median, keys, operation_time))

    results = table_new(attempts, 0)
    for _ = 1, attempts do
        table.insert(results, stopwatch.measure(function() test2(iterations) end, {format = 'ms'}))
    end
    table.sort(results)
    median = results[math.floor(attempts / 2)]
    operation_time = (median / iterations) * 1e6
    print(string.format("test2 - t = {} median: %.2fms, %d insertions - %.2fns/op", median, keys, operation_time))

    results = table_new(attempts, 0)
    for _ = 1, attempts do
        table.insert(results, stopwatch.measure(function() test3(iterations) end, {format = 'ms'}))
    end
    table.sort(results)
    median = results[math.floor(attempts / 2)]
    operation_time = (median / iterations) * 1e6
    print(string.format("test3 - t = table.new(0, 0) median: %.2fms, %d insertions - %.2fns/op", median, keys, operation_time))
end

bench(1e5)

results.txt

$ tarantool bench.lua
test1 - t = table.new(0, 100) median: 12.05ms, 100 insertions - 120.50ns/op
test2 - t = {} median: 122.67ms, 100 insertions - 1226.68ns/op
test3 - t = table.new(0, 0) median: 124.47ms, 100 insertions - 1244.68ns/op