Check out the new USENIX Web site. next up previous
Next: Applications Up: Performance Evaluation Previous: MultiView Limitations

   
Basic Costs in MILLIPAGE

A major goal in MILLIPAGE's design was to minimize the DSM protocol time by implementing a thin protocol layer. Table 1 shows the measured times (costs) of some basic operations that are part of the DSM protocols in MILLIPAGE.


 
Table 1: Cost of basic operations in MILLIPAGE.
operation $\mu{}s$
access fault 26
get protection 7
set protection 12
header message send/recv (32 bytes) 12
a data message send/recv (0.5 KB) 22
a data message send/recv (1 KB) 34
a data message send/recv (4 KB) 90
minipage translation (MPT lookup) 7
 

The time it takes to bring in a page for reading in MILLIPAGE is 204 $\mu{}s$ for minipages of size 128 byte, and 314 $\mu{}s$ for minipages of size 4 KB. The difference in arrival times for a minipage request arriving in a single hop as opposed to two hops was slight. The time it takes to bring in a page for writing in MILLIPAGE is 212-366 $\mu{}s$ for 128 bytes minipages, and 327-480 $\mu{}s$ for 4KB minipages. These times vary according to the number of read copies that should be invalidated prior to serving the write request.

A barrier between 1 to 8 hosts takes 59-153 $\mu{}s$ (linearly in the number of hosts) and a lock followed by an unlock operation takes 67-80 usec.

In addition, we measured diff creation time in our setting. Our measurements show that a run-length diff operation (as described in [5]) for 4KB page takes 250 $\mu{}s$ and decreases linearly with the size of the page. Obviously, this time is not negligible, and would have dominated the overhead if it were required in the DSM protocol.


next up previous
Next: Applications Up: Performance Evaluation Previous: MultiView Limitations
Ayal Itzkovitz and Assaf Schuster, The Technion