Missing Writes with 99/1 Read/Write ratio

classic Classic list List threaded Threaded
5 messages Options
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Missing Writes with 99/1 Read/Write ratio

Mark_Lindholm

I have attached a simple script that repeats a 4K Sequential 99/1 Read/Write with incrementing Q depth.

 

I have also attached the test output, which shows that the writes are sometimes missing.

 

If I change the Read/Write ratio to anything other than 99/1 (or 1/99 the reads will sometimes stop).

 

Any idea why this does not work correctly?

 

Mark.


------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user

Ramp0_Run10_Phoenix.xlsx (798K) Download Attachment
4K_99_1_512balign_Ramp0_Run60_080113.icf (11K) Download Attachment
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Missing Writes with 99/1 Read/Write ratio

Vedran Degoricija
Hi Mark,
 
This is interesting. How long do you typically run each test for?
 
I suspect the problem is with the fairness of the indexing algorithm into the access_grid (see IOAccess.*) that represent the read/write mix.  With one entry being either a write or a read, it has really low chances of getting hit. Unless something is broken,  we'd need to look into a fairer algorithm or you run for a longer time.
 
Ved

From: "[hidden email]" <[hidden email]>
To: [hidden email]
Sent: Tuesday, August 6, 2013 3:41 PM
Subject: [Iometer-user] Missing Writes with 99/1 Read/Write ratio

I have attached a simple script that repeats a 4K Sequential 99/1 Read/Write with incrementing Q depth.
 
I have also attached the test output, which shows that the writes are sometimes missing.
 
If I change the Read/Write ratio to anything other than 99/1 (or 1/99 the reads will sometimes stop).
 
Any idea why this does not work correctly?
 
Mark.

------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user



------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Missing Writes with 99/1 Read/Write ratio

Mark_Lindholm

In the example I attached I was running for 60 seconds for each Q depth.

 

So for some of the workloads/Qdepth  I was not getting a single write in 60 seconds of run.

 

I put a bus analyzer on when I was running and it confirmed the 1%  writes, were sometimes not there.

 

I also tried other ratios, like 98/2, 97,3 and they worked but they would not give me the exact ratio asked for (according to the bus analyzer).

 

So if I asked for 98/2, I  might get 99/1 or 97/3.  Following the same logic if I ask for 99/1 I sometimes get 2% writes and sometimes get 0%?

 

Mark.

 

From: Vedran Degoricija [mailto:[hidden email]]
Sent: Tuesday, August 20, 2013 11:28 PM
To: Lindholm, Mark; [hidden email]
Subject: Re: [Iometer-user] Missing Writes with 99/1 Read/Write ratio

 

Hi Mark,

 

This is interesting. How long do you typically run each test for?

 

I suspect the problem is with the fairness of the indexing algorithm into the access_grid (see IOAccess.*) that represent the read/write mix.  With one entry being either a write or a read, it has really low chances of getting hit. Unless something is broken,  we'd need to look into a fairer algorithm or you run for a longer time.

 

Ved

 

From: "[hidden email]" <[hidden email]>
To: [hidden email]
Sent: Tuesday, August 6, 2013 3:41 PM
Subject: [Iometer-user] Missing Writes with 99/1 Read/Write ratio

 

I have attached a simple script that repeats a 4K Sequential 99/1 Read/Write with incrementing Q depth.

 

I have also attached the test output, which shows that the writes are sometimes missing.

 

If I change the Read/Write ratio to anything other than 99/1 (or 1/99 the reads will sometimes stop).

 

Any idea why this does not work correctly?

 

Mark.


------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user


------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Missing Writes with 99/1 Read/Write ratio

Vedran Degoricija
Hi Mark,
 
Yes, unfortunately, this consistent. The higher the R/W difference (excepting 100/0), the lower the chances of hitting the smaller valued I/O. This should be addressed. :(
 
I am curious, under what OS was Dynamo executing?
 
Thanks,
Ved

From: "[hidden email]" <[hidden email]>
To: [hidden email]; [hidden email]
Sent: Wednesday, August 21, 2013 7:06 AM
Subject: RE: [Iometer-user] Missing Writes with 99/1 Read/Write ratio

In the example I attached I was running for 60 seconds for each Q depth.
 
So for some of the workloads/Qdepth  I was not getting a single write in 60 seconds of run.
 
I put a bus analyzer on when I was running and it confirmed the 1%  writes, were sometimes not there.
 
I also tried other ratios, like 98/2, 97,3 and they worked but they would not give me the exact ratio asked for (according to the bus analyzer).
 
So if I asked for 98/2, I  might get 99/1 or 97/3.  Following the same logic if I ask for 99/1 I sometimes get 2% writes and sometimes get 0%?
 
Mark.
 
From: Vedran Degoricija [mailto:[hidden email]]
Sent: Tuesday, August 20, 2013 11:28 PM
To: Lindholm, Mark; [hidden email]
Subject: Re: [Iometer-user] Missing Writes with 99/1 Read/Write ratio
 
Hi Mark,
 
This is interesting. How long do you typically run each test for?
 
I suspect the problem is with the fairness of the indexing algorithm into the access_grid (see IOAccess.*) that represent the read/write mix.  With one entry being either a write or a read, it has really low chances of getting hit. Unless something is broken,  we'd need to look into a fairer algorithm or you run for a longer time.
 
Ved
 
From: "[hidden email]" <[hidden email]>
To: [hidden email]
Sent: Tuesday, August 6, 2013 3:41 PM
Subject: [Iometer-user] Missing Writes with 99/1 Read/Write ratio
 
I have attached a simple script that repeats a 4K Sequential 99/1 Read/Write with incrementing Q depth.
 
I have also attached the test output, which shows that the writes are sometimes missing.
 
If I change the Read/Write ratio to anything other than 99/1 (or 1/99 the reads will sometimes stop).
 
Any idea why this does not work correctly?
 
Mark.

------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user




------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user
Reply | Threaded
Open this post in threaded view
|  
Report Content as Inappropriate

Re: Missing Writes with 99/1 Read/Write ratio

Mark_Lindholm

I am running Windows Server 2008 R2.  Running Dynamo directly (no network).

OMETER 2006.07.27

 

 

Mark.

 

From: Vedran Degoricija [mailto:[hidden email]]
Sent: Wednesday, August 21, 2013 9:17 PM
To: Lindholm, Mark; [hidden email]
Subject: Re: [Iometer-user] Missing Writes with 99/1 Read/Write ratio

 

Hi Mark,

 

Yes, unfortunately, this consistent. The higher the R/W difference (excepting 100/0), the lower the chances of hitting the smaller valued I/O. This should be addressed. :(

 

I am curious, under what OS was Dynamo executing?

 

Thanks,
Ved

 

From: "[hidden email]" <[hidden email]>
To: [hidden email]; [hidden email]
Sent: Wednesday, August 21, 2013 7:06 AM
Subject: RE: [Iometer-user] Missing Writes with 99/1 Read/Write ratio

 

In the example I attached I was running for 60 seconds for each Q depth.

 

So for some of the workloads/Qdepth  I was not getting a single write in 60 seconds of run.

 

I put a bus analyzer on when I was running and it confirmed the 1%  writes, were sometimes not there.

 

I also tried other ratios, like 98/2, 97,3 and they worked but they would not give me the exact ratio asked for (according to the bus analyzer).

 

So if I asked for 98/2, I  might get 99/1 or 97/3.  Following the same logic if I ask for 99/1 I sometimes get 2% writes and sometimes get 0%?

 

Mark.

 

From: Vedran Degoricija [[hidden email]]
Sent: Tuesday, August 20, 2013 11:28 PM
To: Lindholm, Mark; [hidden email]
Subject: Re: [Iometer-user] Missing Writes with 99/1 Read/Write ratio

 

Hi Mark,

 

This is interesting. How long do you typically run each test for?

 

I suspect the problem is with the fairness of the indexing algorithm into the access_grid (see IOAccess.*) that represent the read/write mix.  With one entry being either a write or a read, it has really low chances of getting hit. Unless something is broken,  we'd need to look into a fairer algorithm or you run for a longer time.

 

Ved

 

From: "[hidden email]" <[hidden email]>
To: [hidden email]
Sent: Tuesday, August 6, 2013 3:41 PM
Subject: [Iometer-user] Missing Writes with 99/1 Read/Write ratio

 

I have attached a simple script that repeats a 4K Sequential 99/1 Read/Write with incrementing Q depth.

 

I have also attached the test output, which shows that the writes are sometimes missing.

 

If I change the Read/Write ratio to anything other than 99/1 (or 1/99 the reads will sometimes stop).

 

Any idea why this does not work correctly?

 

Mark.


------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user

 


------------------------------------------------------------------------------
Introducing Performance Central, a new site from SourceForge and
AppDynamics. Performance Central is your source for news, insights,
analysis and resources for efficient Application Performance Management.
Visit us today!
http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
_______________________________________________
Iometer-user mailing list
[hidden email]
https://lists.sourceforge.net/lists/listinfo/iometer-user
Loading...