Interestingly enough I was given this problem (well, numerically stable online variance, but close enough) a while ago as a take-home job interview and I ended up researching this quite a bit.

It has been a while, but if I remember correctly pairwise sum was in practice as good or better as Kahan summation (and both more stable than Walford's) but faster.

But by far the fastest and still very stable solution was to use the legacy 80 bit x87 fpu as accumulator. Which is not surprising as Kahan explicitly added 80 bit precision to x87 for intermediate computations.