CSC/ECE 506 Fall 2007/wiki3 7 qaz

Wiki: True and false sharing. In Lectures 9 and 10, we covered performance results for true- and false-sharing misses. The results showed that some applications experienced degradation due to false sharing, and that this problem was greater with larger cache lines. But these data are at least 9 years old, and for multiprocessors that are smaller than those in use today. Comb the ACM Digital Library, IEEE Xplore, and the Web for more up-to-date results. What strategies have proven successful in combating false sharing? Is there any research into ways of diminishing true-sharing misses, e.g., by locating communicating processes on the same processor? Wouldn't this diminish parallelism and thus hurt performance?

False Sharing Miss

In multiprocessor system, it is important to ensure data coherence across all processors. Vendors like Intel uses MESI protocol to ensure cache coherence. When the program is loaded on to the cache for the first time, MESI requires to put this in Exclusive state. This data in the processors cache will go to shared state when another processor requests the same portion of program. For all subsequent stores by any one processor will cause its state to change from shared to modified and it will invalidate the corresponding cache content of other processors. Figure 1 demonstrates how two distinct variables that are placed adjacent to each other in system memory can be loaded on two or more processors cache line, causing the processor to mark the whole line as shared and invalidate the line for each load/store.

Fasle sharing is caused because of multiple processes/threads sharing same address space, and it is very popular to occur on multiprocessor system. MESI protocol, which is basically an invalidate based coherence protocol, will invalidate shared cache lines on different processors for each load/store of an element in a shared cache line. Figure 2 demostrates this. Even though CPU 0 is accessing different word in the cache line than what CPU 1 is trying to, the entire line in CPU 1 cache will be invalidated. A false sharing miss will occur when CPU 1 again tries to access the same word.

Block Size and False Sharing Miss

The cache performance of shared data in a multiprocessor environment gets affected more because of misses due to shared data than because of any other types of misses. It has been obeserved that shared data is responsible for majority of cache misses, and there miss rate is less predictable. Generally miss rate in uniprocessor goes down with increase in block size, but in multi-processor it goes up with increase in block size (more often, but this is not a general case). The misses because of shared data is also know as coherence miss, and is further of two types. Miss due to false sharing and miss due to true sharing.

Coherence miss that are casued because of false sharing is a subject of speculation. As mentioned above, coherence miss increases with increase in block size is because of higher false sharing of data with larger blocks. Even though we have generalised this, but it sometimes depends on the application that is being run on system. Below it is show contribution of each types of misses for different application.

The figure shows 6 bars in each set. Each bar of a set corresponds to a scenario of cache. There are 5 different kinds scenarios:

Baseline : Basic invalidation based MESI protocol.

UFS  : Updatebased false sharing.

UFS-P  : Measures the potential of UFS with perfect knowledge. This allows us to avoid reading followed by an upgrade (S->M).

MSFS  : Message passing stochastic false sharing.

PSFS  : Program structure stochastic false sharing.

Figure shows that false sharing increses with the increase in block size. Also, the steps of increase is very small in case of OLTP, and in fact the false sharing decreases with block size 512B. Hence, in general the false sharing increase with increase in block size but its not always true.

Strategies to reduce False Sharing Miss

Many of the proposed strategies to improve miss rate because of false sharing, revolve around data transformation. J. Torrellas, M. Lam, and J. Hennessey proposed methods like array padding and block alignment to reduce false sharing. S. Eggers and T. Jeremiassen proposed a method to reduce false sharing either by grouping data that is accessed by the same processor or separate individual data items that are shared. Basic techniques proposed that fall under data transformation can be listed below:

Changing loop structures : Transform program loops, e.g., by blocking, alignment, or peeling, so that iterations in a parallel loop access disjoint cache lines.

Changing data structures : Change the layout of data structures, e.g., by array alignment and padding. Array alignment is the insertion of dummy space so as to change the starting address of an array variable. Array padding is an increase in the allocated dimension size of an array variable.

Copying data : Copy the data to be updated by the loop into a temporary data structure that does not exhibit false sharing and is well suited to the data access patterns in the loop. After the parallel loop completes execution, the temporary data structure is copied back to the original structure. The copy back may exhibit false sharing, however.

Changing schedule parameters : Schedule the loop iterations so that concurrently executed iterations access disjoint cache lines.

Some of the transformations help improve spatial locality along with reducing false sharing, others may adversely affect locality. Hence there has to be a trade off between spatial locality and false sharing. Kandemir, M. Choudhary, A. Ramanujam and J. Banerjee proposed unified compilation framework in which focus was more on structured codes which demostrated how spatial locality and false sharing can be treated in an optimizing compiler framework.

framework