My favorites | Sign in
Project Home Downloads Wiki Issues Source
READ-ONLY: This project has been archived. For more information see this post.
Search
for
  Advanced search   Search tips   Subscriptions
Issue 60: Definition of ONLINE
2 people starred this issue and may be notified of changes. Back to list
Status:  Done
Owner:  postri...@gmail.com
Closed:  Feb 2009


 
Reported by postri...@gmail.com, Feb 2, 2009
Hello everybody,

I have still not got the idea of "online" HSM properly.

Can SKS and RV write a justification of "How OHSM is online when writes and
reads hanged up"?

Feb 3, 2009
Project Member #1 sandeepksinha
Ok,

So, Online here means as Online in the World of HSM.
If you look at the other types of HSM, some of them is not online. Meaning the
relocation of data is done at the cost of name space changes, as a new inode is
created for it.
for example, if they use CP to copy a complete mountpoint from one place to the other.

Now how OHSM is better than any other online HSM,

For example,
1. Adhoc - Stubs causes lot of time to move the data to a slower disk. And whenver it
is accessed it first brings the data back to the original disk and serves the request. 

We don't use stubs making it faster.


Now coming to what exactly happens when reads and write are going on?

Remember that even if the file is not accessible for small period of time, only if
read/writes are going on that file.

Also, this time is very small, and can be neglected as our major focus is to have HSM
framework, which is a long term goal. Also, none of the applications break, its
completely transparent to any users.

If they say there is timeouts,
Then obviously, you have the methods of try/catch/throw to retry the same.

Also, we lock the file for some time, but don't forget that all further i/o's on it
will be faster after that.

One more thing,
There can two kinds of relocation,
1. High to low speed tier -- This happens only when a file has not been referred from
a long time. And the possibility of this file being open will be 0.1%.

2. Low to High,
Neways the already ongoing i/o would be slower as we are serving it from a slower disk.
So, if the amount of data to be read is 100K.
And the speed on slower disk is 1KB/sec.
And on a faster disk is 5K/Sec.

Without relocation(OHSM) the application would take 100Sec.

So, if we read 20 KB from slower disk --  20Sec
And to relocate it takes 5 Sec.
remaining 80KB will be from faster disk ---  14 Sec

So, still the overall time would be 39 Sec as compared to 100 Sec.

So, dont just look on how much time we are taking to relocate, one should only focus
on how much overall time is required without OHSM and With OHSM.


I hope this will clarify everything.


Keep 2 slides at the end to show if you feel such situations arrive, else don't show it.


Makes Sense ?????


Feb 3, 2009
Project Member #2 bharati....@gmail.com
Can you clarify a little more on namespace changes? As in, how would you define a
namespace in this context?

Any reason why the stubs used in HSM slow the data path? 

Curiosity didn't kill the cat. It made it wiser :)
Feb 3, 2009
Project Member #3 sandeepksinha
Namespace here means...
in your code if you have written fopen($MNTPT/a.c);

Things like this is written so that when you mount the same device somewhere else and
compile,
automatically it should compile with the current mountpoint.
But once you change the mount point by copying the data to other mountpoint.
Applications will break,

It basically means change in the path name.

The stubs slows down means...
Its juts like the idea where you create a new inode and relocate the data there.


What you do to the original inode is that "All read/write" will first restore the
data blocks and then perform the i/o.
Making it slower.

So stubs means you add a dummy code which does the actual work along with what you
want to,

 
Feb 3, 2009
Project Member #4 sandeepksinha
(No comment was entered for this change.)
Summary: Definition of ONLINE
Feb 3, 2009
#5 postri...@gmail.com
The explanation is quite good and satisfying. RV take a note of this. This is your
job to convince the public that we are ONLINE.
Feb 3, 2009
#6 postri...@gmail.com
(No comment was entered for this change.)
Status: Done

Powered by Google Project Hosting