The Bandwidth

Figures (a) and (b) show the bandwidth obtained between two nodes of the CM-5 for

• the native message passing system CMMD_send_block()-CMMD_receive_block()
• pvm_psend()-pvm_precv()
• pvm_send()-pvm_recv() with the PvmDataRaw format
• pvm_send()-pvm_recv() with the PvmDataInPlace format

 
Figure: Bandwidth on the CM-5: PVM3 - CMMD  

The first point to notice in Figure (a) is that, as expected, the native message-passing library is the most efficient, with an asymptotic bandwidth of 8.06 Mbytes/sec. Nevertheless, the pvm_psend() - pvm_precv() bandwidth is fairly close to the performance of the native system. In fact, pvm_psend() is built on top of CMMD_send_async(), a configuration that explains the good performance (see section ).

We also see that pvm_send()-pvm_recv() with PvmDataInPlace is much less efficient than pvm_psend() - pvm_precv(). Two factors explain this inefficiency:

• Unlike pvm_psend()-pvm_precv(), it involves a real data unpacking on the receiving end (see Figure ).
• It is built on top of CMMD_send_noblock() which we showed much less efficient than CMMD_send_async() because of an extra data copy on the sending end (see section ).

Of course, pvm_send()-pvm_recv() with PvmDataRaw is even less efficient. It is built on top of CMMD_send_async but involves an actual data packing-unpacking.