The Bandwidth

In Figures (a) and (b) we show the bandwidth obtained for

• mpc_bsend()-mpc_brecv()
• pvm_psend()-pvm_precv()
• PVMe : PvmDataInPlace Interrupt/No Interrupt
• PVMe : PvmDataRaw Interrupt/No Interrupt
• pvm_send()-pvm_recv() with PvmDataInPlace
• pvm_send()-pvm_recv() with PvmDataRaw

 
Figure: Bandwidth on the IBM SP-2: PVM3 - PVMe - MPL 

In Figure (a), the measures for PVMe using PvmDataInPlace are the same regardless of whatever execution mode is used. The measures for PVMe using PvmDataRaw in the ``No Interrupt" mode are the same as the measures for PVMe using PvmDataInPlace.

In Figure (b), the measures for PVMe using PvmDataRaw are exactly the same as the measures PVMe using PvmDataInPlace.

In Figure (a), we see that pvm_psend - pvm_precv is only slightly less efficient than mpc_bsend()- mpc_brecv(), which is of course the most efficient. As on the CM-5 and the Paragon, PvmDataInPlace is better than PvmDataRaw.

As with the Paragon, the little blip in the middle of the pvm_send()-pvm_recv() curve corresponds to the system page size.