According to the documentation for ParallelDo (look under Possible Issues), "Side effects are local to each parallel kernel". The example there is analogous to yours. You can share a variable across kernels with SetSharedVariable.

There is overhead in setting up parallel computations, and I would imagine that sharing a variable adds even more overhead. And so, in your example, performance is actually worse with parallelization than without:

Count[t, _?PrimeQ]
(* Runs in 3 seconds on my machine *)

Do[If[PrimeQ[t[[i]]], howmany++], {i, 10^7}]
(* Runs in 7 seconds *)

ParallelDo[If[PrimeQ[t[[i]]], howmany++], {i, 10^7}]
(* Runs in 9 seconds and produces the wrong answer because howmany isn't shared *)

I don't know if this example is representative of your real scenario, but my suggestions would be:

• Use the built-in functions designed for list processing first before you try to optimize "by hand".
• Avoid Do and related constructs unless the scenario actually demands it. Not only are these constructs generally less performant, but they aren't really idiomatic of the Wolfram Language.
• To provide any benefit, parallel constructs require special care to set up properly, so use them only when the overhead is worth it.