Hi All,
Currently we are preparing the final result row wise, as number of columns present in project list(80 columns) is high mainly measure column or no dictionary column there are lots of cpu cache invalidation is happening and this is resulting to slower the query performance. *I can think of two solutions for this problem.* *Solution 1*. Fill column data vertically, currently it is horizontally(It may not solve all the problem) *Solution 2*. Use code generation for result preparation. This is an initially idea. -Regards Kumar Vishal
kumar vishal
|
Hi Vishal,
Which part of the preparation are you considering? The column stitching in the executor side? Regards, Jacky > 在 2016年10月12日,下午9:24,Kumar Vishal <[hidden email]> 写道: > > Hi All, > Currently we are preparing the final result row wise, as number of columns > present in project list(80 columns) is high mainly measure column or no > dictionary column there are lots of cpu cache invalidation is happening and > this is resulting to slower the query performance. > > *I can think of two solutions for this problem.* > *Solution 1*. Fill column data vertically, currently it is horizontally(It > may not solve all the problem) > *Solution 2*. Use code generation for result preparation. > > This is an initially idea. > > -Regards > Kumar Vishal |
Hi Jacky,
Yes result preparation in exeutor side. -Regards Kumar Vishal On Wed, Oct 12, 2016 at 9:33 PM, Jacky Li <[hidden email]> wrote: > Hi Vishal, > > Which part of the preparation are you considering? The column stitching in > the executor side? > > Regards, > Jacky > > > 在 2016年10月12日,下午9:24,Kumar Vishal <[hidden email]> 写道: > > > > Hi All, > > Currently we are preparing the final result row wise, as number of > columns > > present in project list(80 columns) is high mainly measure column or no > > dictionary column there are lots of cpu cache invalidation is happening > and > > this is resulting to slower the query performance. > > > > *I can think of two solutions for this problem.* > > *Solution 1*. Fill column data vertically, currently it is > horizontally(It > > may not solve all the problem) > > *Solution 2*. Use code generation for result preparation. > > > > This is an initially idea. > > > > -Regards > > Kumar Vishal > > > >
kumar vishal
|
Hi Vishal,
In general, it is good idea to have a cache efficient algorithm. For solution-1 : how do you want to handle variable length columns and nulls? may be you will have to maintain variable length columns separately and use offsets ? For solution 2: code generation may be more efficient solution. We should find out all other places in executor that can benefit from code generation apart from row formation. BTW, any specific code generation library you have mind? Best Regards, Aniket On Wed, Oct 12, 2016 at 10:02 AM, Kumar Vishal <[hidden email]> wrote: > Hi Jacky, > Yes result preparation in exeutor side. > > -Regards > Kumar Vishal > > On Wed, Oct 12, 2016 at 9:33 PM, Jacky Li <[hidden email]> wrote: > > > Hi Vishal, > > > > Which part of the preparation are you considering? The column stitching > in > > the executor side? > > > > Regards, > > Jacky > > > > > 在 2016年10月12日,下午9:24,Kumar Vishal <[hidden email]> 写道: > > > > > > Hi All, > > > Currently we are preparing the final result row wise, as number of > > columns > > > present in project list(80 columns) is high mainly measure column or no > > > dictionary column there are lots of cpu cache invalidation is happening > > and > > > this is resulting to slower the query performance. > > > > > > *I can think of two solutions for this problem.* > > > *Solution 1*. Fill column data vertically, currently it is > > horizontally(It > > > may not solve all the problem) > > > *Solution 2*. Use code generation for result preparation. > > > > > > This is an initially idea. > > > > > > -Regards > > > Kumar Vishal > > > > > > > > > |
Hi Vishal,
I think, we need both solution 1 & 2 Solution1 may need re-desiging several parts of Carbon's query process starting from scanner, aggregator to result preparation. This can help avoid the frequent cache invalidation. In Solution2 code generation will not solve the frequent cache invalidation problem. However, It will surely help to improve the performance by having specialised code instead of executing generalised code. Especially as we support several data types and our code is generalised for that. Code generation will help to improve performance. Regards Vimal On Thu, Oct 13, 2016 at 3:02 AM, Aniket Adnaik <[hidden email]> wrote: > Hi Vishal, > > In general, it is good idea to have a cache efficient algorithm. > > For solution-1 : how do you want to handle variable length columns and > nulls? may be you will have to maintain variable length columns separately > and use offsets ? > > For solution 2: code generation may be more efficient solution. We should > find out all other places in executor that can benefit from code generation > apart from row formation. BTW, any specific code generation library you > have mind? > > Best Regards, > Aniket > > On Wed, Oct 12, 2016 at 10:02 AM, Kumar Vishal <[hidden email]> > wrote: > > > Hi Jacky, > > Yes result preparation in exeutor side. > > > > -Regards > > Kumar Vishal > > > > On Wed, Oct 12, 2016 at 9:33 PM, Jacky Li <[hidden email]> wrote: > > > > > Hi Vishal, > > > > > > Which part of the preparation are you considering? The column stitching > > in > > > the executor side? > > > > > > Regards, > > > Jacky > > > > > > > 在 2016年10月12日,下午9:24,Kumar Vishal <[hidden email]> 写道: > > > > > > > > Hi All, > > > > Currently we are preparing the final result row wise, as number of > > > columns > > > > present in project list(80 columns) is high mainly measure column or > no > > > > dictionary column there are lots of cpu cache invalidation is > happening > > > and > > > > this is resulting to slower the query performance. > > > > > > > > *I can think of two solutions for this problem.* > > > > *Solution 1*. Fill column data vertically, currently it is > > > horizontally(It > > > > may not solve all the problem) > > > > *Solution 2*. Use code generation for result preparation. > > > > > > > > This is an initially idea. > > > > > > > > -Regards > > > > Kumar Vishal > > > > > > > > > > > > > > > |
Free forum by Nabble | Edit this page |