Line: 1 to 1  

 
Added:  
> >  For our more recent work on the Sparse MatrixVector Multiplication check: CSX.  
Sparse MatrixVector MultiplicationSparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm, which is further aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented 
Line: 1 to 1  

Sparse MatrixVector Multiplication  
Line: 11 to 11  
 
Added:  
> > 
 

Line: 1 to 1  

Sparse MatrixVector Multiplication  
Line: 8 to 8  
Another important optimization that has been widely applied to SPMV is blocking. While such optimization was initially intended to increase register reuse and the locality of references to the input vector, the main advantage of blocking is that it can significantly reduce the working set of the algorithm, thus leading to less memory traffic, meaning better performance for the case of the memorybound SPMV. After attacking the memory intensiveness of the algorithm, the computational part of the kernel becomes more significant as a result. This allows for computation optimizations, such as vectorization, which can add a further performance boost to the blocked kernel. However, preliminary experiments have shown that the performance among different block shapes, which lead to the same working set, can have significantly varying performance, especially in the case of vectorization, due to a number of architectural issues raised from modern commodity microprocessors. Considering the multitude of blocked storage formats available for sparse matrices and the fact that block shapes can heavily affect the performance of SPMV, a performance model that could accurately predict the performance of a blocked storage format and then guide through the proper selection of both the right method and the right block shape would be desirable. Towards this direction, we have developed a performance model that accounts for both the memory and computation times of the blocked kernels. Our model is capable to better approximate the real execution time and lead to a more fit selection of method and block shape than earlier models, which solely based their prediction to memory traffic.
Publications  
Changed:  
< < 
 
> > 
 
Diploma Thesis

Line: 1 to 1  

Sparse MatrixVector Multiplication  
Line: 8 to 8  
Another important optimization that has been widely applied to SPMV is blocking. While such optimization was initially intended to increase register reuse and the locality of references to the input vector, the main advantage of blocking is that it can significantly reduce the working set of the algorithm, thus leading to less memory traffic, meaning better performance for the case of the memorybound SPMV. After attacking the memory intensiveness of the algorithm, the computational part of the kernel becomes more significant as a result. This allows for computation optimizations, such as vectorization, which can add a further performance boost to the blocked kernel. However, preliminary experiments have shown that the performance among different block shapes, which lead to the same working set, can have significantly varying performance, especially in the case of vectorization, due to a number of architectural issues raised from modern commodity microprocessors. Considering the multitude of blocked storage formats available for sparse matrices and the fact that block shapes can heavily affect the performance of SPMV, a performance model that could accurately predict the performance of a blocked storage format and then guide through the proper selection of both the right method and the right block shape would be desirable. Towards this direction, we have developed a performance model that accounts for both the memory and computation times of the blocked kernels. Our model is capable to better approximate the real execution time and lead to a more fit selection of method and block shape than earlier models, which solely based their prediction to memory traffic.
Publications  
Changed:  
< < 
 
> > 
 
 
Added:  
> > 
 
Diploma Thesis

Line: 1 to 1  

Sparse MatrixVector Multiplication  
Line: 10 to 10  
Publications
 
Added:  
> > 
Diploma Thesis 
Line: 1 to 1  

Sparse MatrixVector MultiplicationSparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm, which is further aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented in CF'08 (pdf)  
Added:  
> >  Another important optimization that has been widely applied to SPMV is blocking. While such optimization was initially intended to increase register reuse and the locality of references to the input vector, the main advantage of blocking is that it can significantly reduce the working set of the algorithm, thus leading to less memory traffic, meaning better performance for the case of the memorybound SPMV. After attacking the memory intensiveness of the algorithm, the computational part of the kernel becomes more significant as a result. This allows for computation optimizations, such as vectorization, which can add a further performance boost to the blocked kernel. However, preliminary experiments have shown that the performance among different block shapes, which lead to the same working set, can have significantly varying performance, especially in the case of vectorization, due to a number of architectural issues raised from modern commodity microprocessors. Considering the multitude of blocked storage formats available for sparse matrices and the fact that block shapes can heavily affect the performance of SPMV, a performance model that could accurately predict the performance of a blocked storage format and then guide through the proper selection of both the right method and the right block shape would be desirable. Towards this direction, we have developed a performance model that accounts for both the memory and computation times of the blocked kernels. Our model is capable to better approximate the real execution time and lead to a more fit selection of method and block shape than earlier models, which solely based their prediction to memory traffic.  
Publications

Line: 1 to 1  

Sparse MatrixVector MultiplicationSparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm, which is further aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented  
Changed:  
< <  in CF'08 (pdf)  
> >  in CF'08 (pdf)  
Publications
 
Changed:  
< < 
 
> > 

Line: 1 to 1  

Sparse MatrixVector MultiplicationSparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm, which is further aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented  
Changed:  
< <  in CF'08 (pdf)  
> >  in CF'08 (pdf)  
Publications
 
Deleted:  
< < 
 
\ No newline at end of file  
Added:  
> > 

Line: 1 to 1  

Sparse MatrixVector Multiplication 
Line: 1 to 1  

Changed:  
< < 
 
> > 
 
Sparse MatrixVector MultiplicationSparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm, which is further aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented 
Line: 1 to 1  

Sparse MatrixVector Multiplication 
Line: 1 to 1  

Sparse MatrixVector Multiplication  
Changed:  
< <  Sparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing poses various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm and aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented  
> >  Sparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm, which is further aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented  
in CF'08 (pdf)
Publications 
Line: 1 to 1  

 
Changed:  
< <  SPMV  
> >  Sparse MatrixVector Multiplication  
Changed:  
< <  Sparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing poses various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm and aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented  
> >  Sparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing poses various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm and aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented  
in CF'08 (pdf)
Publications 
Line: 1 to 1  

SPMV  
Changed:  
< <  Sparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing poses various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" ( pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm and aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented in CF'08 ( pdf)  
> >  Sparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing poses various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" (pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm and aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented in CF'08 (pdf)  
Publications  
Changed:  
< < 
 
> > 

Line: 1 to 1  

SPMV  
Changed:  
< <  Sparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing poses various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm and aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented in Computing Frontiers 2008.  
> >  Sparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing poses various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" ( pdf) published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm and aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented in CF'08 ( pdf)  
Publications
 
Deleted:  
< <   GeorgiosGoumas  06 Mar 2008 
Line: 1 to 1  

 
Changed:  
< <  Sparse MatrixVector Multiplication (SPMV)  
> >  SPMVSparse MatrixVector Multiplication (SPMV) is one of the most important computational kernels, lying at the heart of very effective and popular iterative solution methods like Conjugate Gradient (CG) and Generalized Minimal Residual (GMRES). Sparse matrices are generally stored in special formats (e.g. the Compressed Storage Row format) in order to save memory space and computations. SPMV is a challenging kernel posing poses various problems to the underlying architecture that need to be well understood in order consolidate an effective optimization process. Our work "Understanding the Performance of Sparse MatrixVector Multiplication" published in Euromicro/PDP 2008 performs an extensive experimental evaluation of SPMV that leads to a number of optimization guidelines. The major problem that needs to be addressed is the memory bottleneck caused by the data intensity of the algorithm and aggravated by the extra indexing information kept by the sparse matrix storage format. One approach to deal with this problem is to sacrifice CPU cycles in order to compress the data involved in the kernel and consequently alleviate the memory bus from excessive data transfers. A method to compress the indexing information of the matrix structure and the numerical values of the matrix is presented in Computing Frontiers 2008.
Publications
 
 GeorgiosGoumas  06 Mar 2008 
Line: 1 to 1  

Added:  
> > 
 GeorgiosGoumas  06 Mar 2008 