Does anybody know why there are the dips? As in, it doesn't always increase.
This comment was marked helpful 1 times.
ycp
@selena731, if I was to venture a guess, it would be because when you have different number of processors the way that access to the bus is handled changes in a way that is either better or worse for the program that is graphed here. I would also guess that this would change a lot if you simply ran the program again - the dips might occur in different places. In general, it's just more important to note the direct relationship between processors and time to acquire a lock.
This comment was marked helpful 0 times.
idl
My understanding of why the time increases with the number of processors is that when the processor that holds the lock needs to release it (by gaining access to the bus), all the other processors are stupidly taking turns invalidating and gaining exclusive access, thus causing massive amounts of bus traffic. As a result, the processor that actually needs it has to wait for the arbiter to grant it access; this time increases with the number of processors.
This comment was marked helpful 0 times.
paraU
There is a great paper from MIT CSAIL talking about the spin lock performance for linux scalability: An Analysis of Linux Scalability to Many Cores. It's a very good paper and talks about the same thing in the lecture. But you can see many live examples of its impact on linux scalability.
This comment was marked helpful 1 times.
DunkMaster
@selena731 Another possible reason is that the different topology might be used for different number of processors
Does anybody know why there are the dips? As in, it doesn't always increase.
This comment was marked helpful 1 times.
@selena731, if I was to venture a guess, it would be because when you have different number of processors the way that access to the bus is handled changes in a way that is either better or worse for the program that is graphed here. I would also guess that this would change a lot if you simply ran the program again - the dips might occur in different places. In general, it's just more important to note the direct relationship between processors and time to acquire a lock.
This comment was marked helpful 0 times.
My understanding of why the time increases with the number of processors is that when the processor that holds the lock needs to release it (by gaining access to the bus), all the other processors are stupidly taking turns invalidating and gaining exclusive access, thus causing massive amounts of bus traffic. As a result, the processor that actually needs it has to wait for the arbiter to grant it access; this time increases with the number of processors.
This comment was marked helpful 0 times.
There is a great paper from MIT CSAIL talking about the spin lock performance for linux scalability: An Analysis of Linux Scalability to Many Cores. It's a very good paper and talks about the same thing in the lecture. But you can see many live examples of its impact on linux scalability.
This comment was marked helpful 1 times.
@selena731 Another possible reason is that the different topology might be used for different number of processors
This comment was marked helpful 0 times.