module Xdr_mstring

Managed strings are used in XDR context for constant strings that are stored either as string or as memory (bigarray of char).

A managed string ms is declared in the XDR file as in

      typedef _managed string ms<>;

In the encoded XDR stream there is no difference between strings and managed strings, i.e. the wire representation is identical. Only the Ocaml type differs to which the managed string is mapped. This type is Xdr_mstring.mstring (below).

In the RPC context there is often the problem that the I/O backend would profit from a different string representation than the user of the RPC layer. To bridge this gap, managed strings have been invented. Generally, the user can determine how to represent strings (usually either as an Ocaml string, or as memory), and the I/O backend can request to transform to a different representation when this leads to an improvement (i.e. copy operations can be saved).

Only large managed strings result in a speedup of the program (at least several K).

There are two cases: The encoding case, and the decoding case. In the encoding case the mstring object is created by the user and passed to the RPC library. This happens when a client prepares an argument for calling a remote procedure, or when the server sends a response back to the caller. In the decoding case the client analyzes the response from an RPC call, or the server looks at the arguments of an RPC invocation. The difference here is that in the encoding case user code can directly create mstring objects by calling functions of this module, whereas in the decoding case the RPC library creates the mstring objects.

For simplicity, let us only look at this problem from the perspective of an RPC client.

Encoding. Image a client wants to call an RPC, and one of the arguments is a managed string. This means we finally need an mstring object that can be put into the argument list of the call.

This library supports two string representation specially: The normal Ocaml string type, and Netsys_mem.memory which is actually just a bigarray of char's. There are two factories fac,

• Xdr_mstring.string_based_mstrings, and
• Xdr_mstring.memory_based_mstrings,

and both can be used to create the mstring to pass to the RPC layer. It should be noted that this layer can process the memory representation a bit better. So, if the original data value is a string, the factory for string should be used, and if it is a char bigarray, the factory for memory should be used. Now, the mstring object is created by

• let mstring = fac # create_from_string data pos len copy_flag, or by
• let mstring = fac # create_from_memory data pos len copy_flag.

Of course, if fac is the factory for strings, the create_from_string method works better, and if fac is for memory, the create_from_memory method works better. pos and len can select a substring of data. If copy_flag is false, the mstring object does not copy the data if possible, but just keeps a reference to data until it is accessed; otherwise if copy_flag is true, a copy is made immediately. Of couse, delaying the copy is better, but this requires that data is not modified until the RPC call is completed.

Decoding. Now, the call is done, and the client looks at the result. There is also an mstring object in the result. As noted above, this mstring object was already created by the RPC library (and currently this library prefers string-based objects if not told otherwise). The user code can now access this mstring object with the access methods of the mstring class (see below). As these methods are quite limited, it makes normally only sense to output the mstring contents to a file descriptor.

The user can request a different factory for managed strings. The function Rpc_client.set_mstring_factories can be used for this purpose. (Similar ways exist for managed clients, and for RPC servers.)

Potential. Before introducing managed strings, a clean analysis was done how many copy operations can be avoided by using this technique. Example: The first N bytes of a file are taken as argument of an RPC call. Instead of reading these bytes into a normal Ocaml string, an optimal implementation uses now a memory buffer for this purpose. This gives:

• Old implementation with strings and ocamlnet-2: Data is copied six times from reading it from the file until writing it to the socket.
• New implementation with memory-based mstrings: Data is copied only twice! The first copy reads it from the file into the input buffer (a memory value), and the second copy writes the data into the socket.

Part of the optimization is that Unix.read and Unix.write do a completely avoidable copy of the data which is prevented by switching to Netsys_mem.mem_read and Netsys_mem.mem_write, respectively. The latter two functions exploit an optimization that is only possible when the data is memory-typed.

The possible optimizations for the decoding side of the problem are slightly less impressive, but still worth doing it.