How would you manage a std::vector of structs with

I have a set of structures similar to:

typedef struct {
   int a;
   int b;
} ITEM;

typedef struct {
   int orderID;
   std::vector<ITEM> items;
} ORDER;

typedef struct {
   int orderSetID;
   std::vector<ORDER> Orders;
} ORDER_SET;

The problem is that the number of orders ranges between 100,000 to 10,000,000 and the number of ITEMS in an ORDER ranges from 1 to 500.

The issue is, as I build the ORDER_SET, I don't know how many ORDERs there will be. I do know when I add an ORDER how many ITEMS there will be. Here are some problems:

1) Ideally once I allocate the memory for all of the ORDERs using Orders.resize(), I'd be able to reuse the memory, but it appears that Orders.clear() does delete it all.

2) I start with a somewhat reasonable size for Orders, say Orders.resize(500,000), but the problem is that when I hit large numbers of orders the resize of the Orders vector takes forever because it has to copy all of the ORDER.item vectors.

Seems like a similar issue to storing spare matrixes, the issue being that I also don't know how large the structure is going to be before I create it.

Additional information:
1) Using Visual Studio 2008 2) As posted in the comments below I was able to improve the construction of an ORDER_SET which contains 10000000 Orders in a reasonable time by replacing items with two fields leaving:

typedef struct {
    int orderID;
   ITEM singleItem;
   std::vector<ITEM> *pItems;
} ORDER;

I put all of the pItems into a separate vector which I use for deleting later.

Now the big issue left is that it appears that calling ORDER_SET.Orders.clear() takes quite some time. I'm not sure why though.

Is there a call like clear() which doesn't release the memory, but just sets end to begin? Once the vector has gotten that large, there really isn't any reason to free the memory, since I might need it again.

Second issue can be fixed by using C++11 with move semantics of vectors, so ORDER can be moved, not just deep copied.

ORDER.Orders.clear() takes some time because it must call destructors on all elements.