https://discuss.aerospike.com/t/lua-bytes-speed-an-memory-efficiency/2242

bench.lua

-- Write an unsigned 32 bit integer v to position i in b in little endian encoding.
local function PutUint32(b, i, v)
  b[i]   = v % 256
  b[i+1] = math.floor(v/256) % 256
  b[i+2] = math.floor(v/65536) % 256
  b[i+3] = math.floor(v/16777216) % 256
end

-- Read an unsigned 32 bit integer from position i in b in little endian encoding.
local function Uint32(b, i)
  return b[i] + (b[i+1] * 256) + (b[i+2] * 65536) + (b[i+3] * 16777216)
end


local function bench_write_lua_(r, n, a, b)
  local s = {} -- There is no proper way to preallocate a Lua table.

  for i=1,n,1 do
    PutUint32(s, #s+1, a)
    PutUint32(s, #s+1, b)
  end

  local by = bytes(#s)
  bytes.set_string(by, 1, string.char(unpack(s)))

  r["bench"] = by
  aerospike:update(r)
end

local function bench_write_as_(r, n, a, b)
  local by   = bytes(8*n) -- preallocate, we know how the max is we need.
  local used = 0
  for i=1,n,1 do
    bytes.set_int32_le(by, used+1, a)
    bytes.set_int32_le(by, used+5, b)
    used = used + 8
  end

  bytes.set_size(by, used)

  r["bench"] = by
  aerospike:update(r)
end

local function bench_read_lua_(r)
  local by = r["bench"]
  local s  = {bytes.get_string(by, 1, bytes.size(by)):byte(1, -1)}

  for i=1,#s,8 do
    Uint32(s, i)
    Uint32(s, i+4)
  end
end

local function bench_read_as_(r)
  local by   = r["bench"]
  local size = bytes.size(by)

  for i=1,size,8 do
    bytes.get_int32_le(by, i)
    bytes.get_int32_le(by, i+4)
  end
end


function bench_write_lua(r, a, b)
  for i=1,10000,1 do
    bench_write_lua_(r, 10, a, b)
  end
end

function bench_write_as(r, a, b)
  for i=1,10000,1 do
    bench_write_as_(r, 10, a, b)
  end
end

function bench_read_lua(r)
  for i=1,20000,1 do
    bench_read_lua_(r)
  end
end

function bench_read_as(r)
  for i=1,20000,1 do
    bench_read_as_(r)
  end
end