Looking for more details about "Group varint encoding/decoding" presented in Jeff's slides

Question

I noticed that in Jeff's slides "Challenges in Building Large-Scale Information Retrieval Systems", which can also be downloaded here: http://research.google.com/people/jeff/WSDM09-keynote.pdf, a method of integers compression called "group varint encoding" was mentioned. It was said much faster than 7 bits per byte integer encoding (2X more). I am very interested in this and looking for an implementation of this, or any more details that could help me implement this by myself.

I am not a pro and new to this, and any help is welcome!

Answer 1

That's referring to "variable integer encoding", where the number of bits used to store an integer when serialized is not fixed at 4 bytes. There is a good description of varint in the protocol buffer documentation.

It is used in encoding Google's protocol buffers, and you can browse the protocol buffer source code.

The CodedOutputStream contains the exact encoding function WriteVarint32FallbackToArrayInline:

inline uint8* CodedOutputStream::WriteVarint32FallbackToArrayInline(
    uint32 value, uint8* target) {
  target[0] = static_cast<uint8>(value | 0x80);
  if (value >= (1 << 7)) {
    target[1] = static_cast<uint8>((value >>  7) | 0x80);
    if (value >= (1 << 14)) {
      target[2] = static_cast<uint8>((value >> 14) | 0x80);
      if (value >= (1 << 21)) {
        target[3] = static_cast<uint8>((value >> 21) | 0x80);
        if (value >= (1 << 28)) {
          target[4] = static_cast<uint8>(value >> 28);
          return target + 5;
        } else {
          target[3] &= 0x7F;
          return target + 4;
        }
      } else {
        target[2] &= 0x7F;
        return target + 3;
      }
    } else {
      target[1] &= 0x7F;
      return target + 2;
    }
  } else {
    target[0] &= 0x7F;
    return target + 1;
  }
}

The cascading ifs will only add additional bytes onto the end of the target array if the magnitude of value warrants those extra bytes. The 0x80 masks the byte being written, and the value is shifted down. From what I can tell, the 0x7f mask causes it to signify the "last byte of encoding". (When OR'ing 0x80, the highest bit will always be 1, then the last byte clears the highest bit (by AND'ing 0x7f). So, when reading varints you read until you get a byte with a zero in the highest bit.

I just realized you asked about "Group VarInt encoding" specifically. Sorry, that code was about basic VarInt encoding (still faster than 7-bit). The basic idea looks to be similar. Unfortunately, it's not what's being used to store 64bit numbers in protocol buffers. I wouldn't be surprised if that code was open sourced somewhere though.

Using the ideas from varint and the diagrams of "Group varint" from the slides, it shouldn't be too too hard to cook up your own :)

Here is another page describing Group VarInt compression, which contains decoding code. Unfortunately they allude to publicly available implementations, but they don't provide references.

void DecodeGroupVarInt(const byte* compressed, int size, uint32_t* uncompressed) {
  const uint32_t MASK[4] = { 0xFF, 0xFFFF, 0xFFFFFF, 0xFFFFFFFF };
  const byte* limit = compressed + size;
  uint32_t current_value = 0;
  while (compressed != limit) {
    const uint32_t selector = *compressed++;
    const uint32_t selector1 = (selector & 3);
    current_value += *((uint32_t*)(compressed)) & MASK[selector1];
    *uncompressed++ = current_value;
    compressed += selector1 + 1;
    const uint32_t selector2 = ((selector >> 2) & 3);
    current_value += *((uint32_t*)(compressed)) & MASK[selector2];
    *uncompressed++ = current_value;
    compressed += selector2 + 1;
    const uint32_t selector3 = ((selector >> 4) & 3);
    current_value += *((uint32_t*)(compressed)) & MASK[selector3];
    *uncompressed++ = current_value;
    compressed += selector3 + 1;
    const uint32_t selector4 = (selector >> 6);
    current_value += *((uint32_t*)(compressed)) & MASK[selector4];
    *uncompressed++ = current_value;
    compressed += selector4 + 1;
  }
}