Using DX8 could be the reason, but I find it strange that UP is twice as fast.

So you're creating the VB in D3DPOOL_DEFAULT, with D3DUSAGE_DYNAMIC|D3DUSAGE_WRITEONLY, and you're locking with D3DLOCK_DISCARD?
And doing the same to the IB? And its still slow?

What does your vertex update loop look like?
If yours is:

for(i=0; i<numVerts; i++)
{
vertex.position.x = newvertex.position.x;
vertex.position.y = newvertex.position.y;
...
}

and theirs is:

memcpy(vertex, newvertex, sizeof(Vertex)*numVerts);

Then that might do it.

How many draws do you do per frame? and how big is your model?

JB