Consistency is a form of reliability

Consistency and reliability are usually considered to be in tension because of the CAP theorem. And when working on a distributed system, it’s normal to get focused on trying to improve service reliability. Sometimes, we can take shortcuts that will increase reliability at the expense of returning stale results more often.

But consistency shouldn’t be seen as opposing reliability. In fact, I think they’re very similar. If your service has 6 9s of reliability (number of successful requests / number of total requests) but only 3 9s of consistency (number of consistent requests / number of successful requests), do you really have 6 9s of reliability? After all, you’re getting some of those extra 9s just by returning a wrong answer.

With my definition of consistency (how many successful requests are consistent), it’s arguably better to prioritize adding more 9s of consistency than reliability.

Why x86 uses xor to zero out a register

I was just reading my colleague Lu Pan’s blog post about C++ exceptions when I saw this quote:

The reason why it uses xor eax eax to set a register to 0 is for efficiency reason. It produces shorter opcode and enables the processor to perform register renaming.

Lu Pan
C++ exception (1) — zero-cost exception handling

I found this rather baffling for a few reasons. First of all, isn’t zeroing out a register a common enough operation that it should be optimized?

Further, why would using xor better for register renaming? Intuitively it seems like it might even be worse since the two operand registers and the result register are all the same.

Instruction size

The obvious way to zero a register is with mov. This instruction takes 7 bytes:

0:  48 c7 c0 00 00 00 00    mov    rax,0x0

It actually includes a full 4 byte immediate value. That wastes quite a bit of space in the instruction.

With xor, the same instruction takes only 3 bytes.

0:  48 31 c0                xor    rax,rax

It pretty much looks like the mov instruction but without the immediate (and a different opcode of course).

Register Renaming

It turns out that Intel has a few dependency breaking idioms in their optimization manual:

Assembly/Compiler Coding Rule 36. (M impact, ML generality) Use dependency-breaking-idiom instructions to set a register to 0, or to break a false dependence chain resulting from re-use of registers.

Section 3.5.1.8 Clearing Registers and Dependency Breaking Idioms
Intel® 64 and IA-32 Architectures Optimization Reference Manual

So actually while it seems strange that xor would be better for register renaming despite referencing the register twice, this particular instruction is explicitly handled by Intel. They themselves recommend it as the best way to zero a register. In fact, the CPU will explicitly skip as many steps as possible on this instruction in order to efficiently zero the register.

New blog to capture my distributed systems learnings

My coworker Lu Pan has a great blog where he writes about his systems and programming learnings. Many times, after we’ve finished debugging a complex issue or having an interesting theoretical discussion, I check his blog and find a post that distills it all down to its key takeaways.

I’m probably not as good as Lu at writing things so succinctly, but his blog inspired me to try it.