Mini filter peformance issues while accessing network files

Hello,

We have hit into serious performance issues while accessing network files with our mini filter driver and we have narrowed it down to 3 filter manager calls that are causing significant performance degradation viz. FltGetFileNameInformation (normalized path), FltQuerySecurityObject, (owner/group and dacl/sacl information), FltQueryInformationFile (basic and standard information)

In order to narrow down the scope for the performance experiments, we tweaked our filter driver to just trap in PostOpCreate and make those 3 calls. We also wrote a simple test application (win32 console application) that just opens and closes the specified file specified number of times and measure the performance in controlled environment.

When we run our test application with our filter driver loaded,

For network files:
FltGetFileNameInformation causes 6x slowness.
FltQuerySecurityObject causes 3x slowness.
FltQueryInformationFile causes 2x slowness.
The cumulative slowness is around 10X.

For local files:
The cumulative slowness is around 2X.

Significant overhead is coming from FltGetFileNameInformation. We are calling it with NameOptions =
FLT_FILE_NAME_NORMALIZED | FLT_FILE_NAME_QUERY_ALWAYS_ALLOW_CACHE_LOOKUP

We found similar threads on the forum in the context of calling it from PreOpCreate. However, we are calling it from PostOpCreate. Further, some threads mention that this was an issue in earlier versions of Windows (e.g. XP). However, we are finding performance degradation in 64-bit Windows 2008 R2 as well

All the three calls are mandatory for functioning of our filter driver. Are there any optimizations that we can look for?

On other note, in our original filter (i.e. not the one tweaked for performance measurements) we cache the name information that we receive from FltGetFileNameInformation call in our stream context (allocated using FltAllocateContext with FLT_STREAM_CONTEXT). This ensures that for subsequent opens, we don’t call the API again if the stream context is around. For local files, we see that Windows keeps the stream context around. However, for network files, we find that stream context is released as soon as the file is closed. Hence, with our test application, we get better numbers if we do open, open, open, close, close, close sequence instead of open,close,open,close,open,close sequence.

Please advice.

Thanks.
-Prasad

What is the protocol server end speaking? what kind of pathnames ?

On Fri, Aug 10, 2012 at 12:15 PM, wrote:
> Hello,
>
> We have hit into serious performance issues while accessing network files with our mini filter driver and we have narrowed it down to 3 filter manager calls that are causing significant performance degradation viz. FltGetFileNameInformation (normalized path), FltQuerySecurityObject, (owner/group and dacl/sacl information), FltQueryInformationFile (basic and standard information)
>
> In order to narrow down the scope for the performance experiments, we tweaked our filter driver to just trap in PostOpCreate and make those 3 calls. We also wrote a simple test application (win32 console application) that just opens and closes the specified file specified number of times and measure the performance in controlled environment.
>
> When we run our test application with our filter driver loaded,
>
> For network files:
> FltGetFileNameInformation causes 6x slowness.
> FltQuerySecurityObject causes 3x slowness.
> FltQueryInformationFile causes 2x slowness.
> The cumulative slowness is around 10X.
>
> For local files:
> The cumulative slowness is around 2X.
>
> Significant overhead is coming from FltGetFileNameInformation. We are calling it with NameOptions =
> FLT_FILE_NAME_NORMALIZED | FLT_FILE_NAME_QUERY_ALWAYS_ALLOW_CACHE_LOOKUP
>
> We found similar threads on the forum in the context of calling it from PreOpCreate. However, we are calling it from PostOpCreate. Further, some threads mention that this was an issue in earlier versions of Windows (e.g. XP). However, we are finding performance degradation in 64-bit Windows 2008 R2 as well
>
> All the three calls are mandatory for functioning of our filter driver. Are there any optimizations that we can look for?
>
> On other note, in our original filter (i.e. not the one tweaked for performance measurements) we cache the name information that we receive from FltGetFileNameInformation call in our stream context (allocated using FltAllocateContext with FLT_STREAM_CONTEXT). This ensures that for subsequent opens, we don’t call the API again if the stream context is around. For local files, we see that Windows keeps the stream context around. However, for network files, we find that stream context is released as soon as the file is closed. Hence, with our test application, we get better numbers if we do open, open, open, close, close, close sequence instead of open,close,open,close,open,close sequence.
>
> Please advice.
>
> Thanks.
> -Prasad
>
>
> —
> NTFSD is sponsored by OSR
>
> For our schedule of debugging and file system seminars visit:
> http://www.osr.com/seminars
>
> To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer

Use OPENED file names if you can - or suffer the performance penalty :slight_smile: There are no workarounds.

xxxxx@vmware.com wrote:

Hello,

We have hit into serious performance issues while accessing network files with our mini filter driver and we have narrowed it down to 3 filter manager calls that are causing significant performance degradation viz. FltGetFileNameInformation (normalized path), FltQuerySecurityObject, (owner/group and dacl/sacl information), FltQueryInformationFile (basic and standard information)

In order to narrow down the scope for the performance experiments, we tweaked our filter driver to just trap in PostOpCreate and make those 3 calls. We also wrote a simple test application (win32 console application) that just opens and closes the specified file specified number of times and measure the performance in controlled environment.

When we run our test application with our filter driver loaded,

For network files:
FltGetFileNameInformation causes 6x slowness.
FltQuerySecurityObject causes 3x slowness.
FltQueryInformationFile causes 2x slowness.
The cumulative slowness is around 10X.

For local files:
The cumulative slowness is around 2X.

Significant overhead is coming from FltGetFileNameInformation. We are calling it with NameOptions =
FLT_FILE_NAME_NORMALIZED | FLT_FILE_NAME_QUERY_ALWAYS_ALLOW_CACHE_LOOKUP

We found similar threads on the forum in the context of calling it from PreOpCreate. However, we are calling it from PostOpCreate. Further, some threads mention that this was an issue in earlier versions of Windows (e.g. XP). However, we are finding performance degradation in 64-bit Windows 2008 R2 as well

All the three calls are mandatory for functioning of our filter driver. Are there any optimizations that we can look for?

On other note, in our original filter (i.e. not the one tweaked for performance measurements) we cache the name information that we receive from FltGetFileNameInformation call in our stream context (allocated using FltAllocateContext with FLT_STREAM_CONTEXT). This ensures that for subsequent opens, we don’t call the API again if the stream context is around. For local files, we see that Windows keeps the stream context around. However, for network files, we find that stream context is released as soon as the file is closed. Hence, with our test application, we get better numbers if we do open, open, open, close, close, close sequence instead of open,close,open,close,open,close sequence.

Please advice.

Thanks.
-Prasad


NTFSD is sponsored by OSR

For our schedule of debugging and file system seminars visit:
http://www.osr.com/seminars

To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer


Kind regards, Dejan (MSN support: xxxxx@alfasp.com)
http://www.alfasp.com
File system audit, security and encryption kits.

Unfortunately Dejan is pretty much right.

The slowdown from FltGetFileNameInformation is because it has to open every component in the path and query for the long name in order to build the normalized name. That is, if the path is \server\share\foo\bar\baz\file.txt, then a call to FltGetFileNameInformation results in:

Open \server\share\foo\bar\baz, QueryDirectory( file.txt, FileNamesInformation )
Open \server\share\foo\bar, QueryDirectory( baz, FileNamesInformation )
Open \server\share\foo, QueryDirectory( bar, FileNamesInformation )
Open \server\share, QueryDirectory( foo, FileNamesInformation )

As you can imagine, this is time consuming on the network and much faster locally.

I’m curious about your cache. Filter Manager already implements a name cache. Calling FltGetFileNameInformation populates this cache the first time around, and draws from the cache on subsequent calls. Filter Manager also takes care of the other work needed to properly maintain this cache (e.g. invalidating it when necessary). Did you implement your own cache in response to seeing poor performance?

In Windows 8 local queries are even faster because Filter Manager is able to use the FileNormalizedNameInformation IRP_MJ_QUERY_INFORMATION class to get the full normalized name in one shot. We weren’t able to get support for that in remote names this time around, but it is something we want to do in the future.

Christian [MSFT]
This posting is provided “AS IS” with no warranties, and confers no rights.

Thank you all for your responses.

@Amritanshu, we are accessing files on a remote Windows share. The paths are of the form \servername\sharename\filepath.

@Dejan, that’s unfortunate to hear. Is “short names not expanded” the only difference between FLT_FILE_NAME_NORMALIZED and FLT_FILE_NAME_OPENED? In that case, can we use some heuristic? e.g. Use FLT_FILE_NAME_OPENED and if the returned path contains any tildes then use FLT_FILE_NAME_NORMALIZED. Agreed, it will cause additional overheads for files accessed using short paths OR if the file genuinely contains tilde character. But at least for files that are accessed using long names, we can optimize.

@Christian, no, we are not maintaining cache specifically for this performance issue. Our filter already maintains some other data in the stream context and one of it is nameinfo returned by FltGetFileNameInformation call. We reference it when we initially create stream context and release it when stream context is destroyed. For network files, we observe that stream context is released as soon as file is closed and hence nameInfo is also released and apparently it is removed from filter manager cache as well and hence for each subsequent open we get a performance overhead. That’s the reason we get better performance for open,open,close,close pattern vs. open, close, open, close pattern for the same file. For local files, stream context remains around and hence the nameinfo as well. Does filter manager also release name information when last reference to the file goes away?

Thanks.
-Prasad

Hello Prasad,

I am not sure how to quantify the 6x performance degradation. At the
risk of stating the obvious, 2k8 onwards SMB2 is used for
2k8 <-> 2k8/Vista or 2k8 <-> Win7/R2, it has a redirector cache this
normally is suppose to speed enumeration and query
info operation for SMB2 clients. Is that enabled for you?

here is a small link for your reference:

http://technet.microsoft.com/en-us/library/ff686200(v=ws.10).aspx

what kind of traffic are you seeing on your netmon/ws trace, too many
disconnects? tree connects?

regards,
Amritanshu.

@Amritanshu, yes, you are right. If both ends (client and server) involve Windows 2008/Windows 7 machines, the degradation is much lower.

I did another experiment *without* our filter loaded. The test application gets the start tick, opens and closes the file on a network share 500 times, gets the end tick and prints the diff between end tick and start tick.

I ran the application on Windows 2008 R2 client. When the network share is hosted on Windows 2003 server, the application took around 850 ticks to complete. However, when the share is hosted on Windows 7 machine, the application took just 47 ticks to complete. That’s significant…

Thanks.
-Prasad

Welcome to the wonderful world of caching (and/or deferral).

Tony
OSR

The file server on Windows 7/Server 2008 takes advantage of a much improved oplock mechanism that allows it to cache more effectively than previous versions.

Does filter manager also release name information when last reference to the file goes away?

Yes.

Christian [MSFT]
This posting is provided “AS IS” with no warranties, and confers no rights.

Handle oplocks.

Further, in Windows 8, handle oplocks are now supported on directories; presumably this should allow caching of even more information (e.g., directory contents).

As an aside, it is interesting that while handle oplocks are supported in Windows 7, they’re actually a bit of a challenge to test as the ifs kit tests don’t actually exercise them.

Tony
OSR

Thank you all for the responses.

Unfortunately, we cannot expect all Windows machines in a network to run Vista/Windows 7(+) to benefit from the optimizations.

Hence, revisiting one of the question that I asked earlier: Is “short names not expanded” the only difference between FLT_FILE_NAME_NORMALIZED and FLT_FILE_NAME_OPENED? In that case, can we use some heuristic? e.g. Use FLT_FILE_NAME_OPENED and if the returned path contains any tildes then use FLT_FILE_NAME_NORMALIZED. Agreed, it will cause additional overheads for files accessed using short paths OR if the file genuinely contains tilde character. But at least for files that are accessed using long names, we can optimize?

Thanks.
-Prasad

Normalized names also resolve mount points.

xxxxx@vmware.com wrote:

Thank you all for the responses.

Unfortunately, we cannot expect all Windows machines in a network to run Vista/Windows 7(+) to benefit from the optimizations.

Hence, revisiting one of the question that I asked earlier: Is “short names not expanded” the only difference between FLT_FILE_NAME_NORMALIZED and FLT_FILE_NAME_OPENED? In that case, can we use some heuristic? e.g. Use FLT_FILE_NAME_OPENED and if the returned path contains any tildes then use FLT_FILE_NAME_NORMALIZED. Agreed, it will cause additional overheads for files accessed using short paths OR if the file genuinely contains tilde character. But at least for files that are accessed using long names, we can optimize?

Thanks.
-Prasad


Kind regards, Dejan (MSN support: xxxxx@alfasp.com)
http://www.alfasp.com
File system audit, security and encryption kits.

I’m not sure how this would work with open by file ID or object ID. Of course, the simplest thing to do would be to simply try it. That’s what Ladislav’s excellent little testing tool is for - so you can open a file in various ways to see what happens, quickly and without extra programming effort.

Tony
OSR

This won’t work because not all short names have the tilde character. You would need to treat any name shorter than 8 characters as potentially a short name.

What do you use the name for?

Thanks,
Alex

Sent from my iPhone

On Aug 16, 2012, at 6:43 AM, Tony Mason wrote:

> I’m not sure how this would work with open by file ID or object ID. Of course, the simplest thing to do would be to simply try it. That’s what Ladislav’s excellent little testing tool is for - so you can open a file in various ways to see what happens, quickly and without extra programming effort.
>
> Tony
> OSR
>
>
> —
> NTFSD is sponsored by OSR
>
> For our schedule of debugging and file system seminars visit:
> http://www.osr.com/seminars
>
> To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer

Are you back?

Mm
On Aug 16, 2012 7:29 AM, “Alex Carp” wrote:

> This won’t work because not all short names have the tilde character. You
> would need to treat any name shorter than 8 characters as potentially a
> short name.
>
> What do you use the name for?
>
> Thanks,
> Alex
>
> Sent from my iPhone
>
> On Aug 16, 2012, at 6:43 AM, Tony Mason wrote:
>
> > I’m not sure how this would work with open by file ID or object ID. Of
> course, the simplest thing to do would be to simply try it. That’s what
> Ladislav’s excellent little testing tool is for - so you can open a file in
> various ways to see what happens, quickly and without extra programming
> effort.
> >
> > Tony
> > OSR
> >
> >
> > —
> > NTFSD is sponsored by OSR
> >
> > For our schedule of debugging and file system seminars visit:
> > http://www.osr.com/seminars
> >
> > To unsubscribe, visit the List Server section of OSR Online at
> http://www.osronline.com/page.cfm?name=ListServer
>
> —
> NTFSD is sponsored by OSR
>
> For our schedule of debugging and file system seminars visit:
> http://www.osr.com/seminars
>
> To unsubscribe, visit the List Server section of OSR Online at
> http://www.osronline.com/page.cfm?name=ListServer
>

I am , but i’m in redmond at plugfest

Sent from my iPhone

On Aug 16, 2012, at 11:10 AM, “Martin O’Brien” wrote:

> Are you back?
>
> Mm
>
> On Aug 16, 2012 7:29 AM, “Alex Carp” wrote:
> This won’t work because not all short names have the tilde character. You would need to treat any name shorter than 8 characters as potentially a short name.
>
> What do you use the name for?
>
> Thanks,
> Alex
>
> Sent from my iPhone
>
> On Aug 16, 2012, at 6:43 AM, Tony Mason wrote:
>
> > I’m not sure how this would work with open by file ID or object ID. Of course, the simplest thing to do would be to simply try it. That’s what Ladislav’s excellent little testing tool is for - so you can open a file in various ways to see what happens, quickly and without extra programming effort.
> >
> > Tony
> > OSR
> >
> >
> > —
> > NTFSD is sponsored by OSR
> >
> > For our schedule of debugging and file system seminars visit:
> > http://www.osr.com/seminars
> >
> > To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer
>
> —
> NTFSD is sponsored by OSR
>
> For our schedule of debugging and file system seminars visit:
> http://www.osr.com/seminars
>
> To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer
> — NTFSD is sponsored by OSR For our schedule of debugging and file system seminars visit: http://www.osr.com/seminars To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer

Outstanding.

mm

From: xxxxx@lists.osr.com [mailto:xxxxx@lists.osr.com] On Behalf Of Alex Carp
Sent: Thursday, August 16, 2012 12:41 PM
To: Windows File Systems Devs Interest List
Cc: Windows File Systems Devs Interest List
Subject: Re: [ntfsd] Mini filter peformance issues while accessing network files

I am , but i’m in redmond at plugfest

Sent from my iPhone

On Aug 16, 2012, at 11:10 AM, “Martin O’Brien” wrote:

Are you back?

Mm

On Aug 16, 2012 7:29 AM, “Alex Carp” wrote:

This won’t work because not all short names have the tilde character. You would need to treat any name shorter than 8 characters as potentially a short name.

What do you use the name for?

Thanks,
Alex

Sent from my iPhone

On Aug 16, 2012, at 6:43 AM, Tony Mason wrote:

> I’m not sure how this would work with open by file ID or object ID. Of course, the simplest thing to do would be to simply try it. That’s what Ladislav’s excellent little testing tool is for - so you can open a file in various ways to see what happens, quickly and without extra programming effort.
>
> Tony
> OSR
>
>
> —
> NTFSD is sponsored by OSR
>
> For our schedule of debugging and file system seminars visit:
> http://www.osr.com/seminars
>
> To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer


NTFSD is sponsored by OSR

For our schedule of debugging and file system seminars visit:
http://www.osr.com/seminars

To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer

— NTFSD is sponsored by OSR For our schedule of debugging and file system seminars visit: http://www.osr.com/seminars To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer


NTFSD is sponsored by OSR

For our schedule of debugging and file system seminars visit:
http://www.osr.com/seminars

To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer

@Dejan, ok, so, we may end up reporting the name without resolving mount points. I am thinking of using the heuristic only for network files.

@Tony, thanks for the pointer. I will give it a try. So you suspect that if the file is opened using file id, opened name may not return us the filename.

@Alex, are you saying that you can have long file names (that don’t fit in 8.3) whose corresponding short filename does not contain tilde character? I looked at http://en.wikipedia.org/wiki/8.3_filename and that doesn’t seem to be the case? We use the name in display and we may re-open the file with that name from our mini-filter.

Thanks.
-Prasad

Yup, that’s what i’m saying. Please don’t forget that it’s possible for a user to set the short name for a file directly as well.

In my opinion if all you do is display the name and use it to re-open then you can just the opened name and not the normalized one. I view the normalized name as necessary only when you’re trying to ‘understand’ the namespace, like when you need to compare paths or possibly do things based on the path ( for path-based policy and such).

Thanks,
Alex.

Sent from my iPhone

On Aug 17, 2012, at 2:35 AM, xxxxx@vmware.com wrote:

@Dejan, ok, so, we may end up reporting the name without resolving mount points. I am thinking of using the heuristic only for network files.

@Tony, thanks for the pointer. I will give it a try. So you suspect that if the file is opened using file id, opened name may not return us the filename.

@Alex, are you saying that you can have long file names (that don’t fit in 8.3) whose corresponding short filename does not contain tilde character? I looked at http://en.wikipedia.org/wiki/8.3_filename and that doesn’t seem to be the case? We use the name in display and we may re-open the file with that name from our mini-filter.

Thanks.
-Prasad


NTFSD is sponsored by OSR

For our schedule of debugging and file system seminars visit:
http://www.osr.com/seminars

To unsubscribe, visit the List Server section of OSR Online at http://www.osronline.com/page.cfm?name=ListServer

@Alex, thanks for pointing this out. However, I guess this is possible only if the end user explicitly sets the short name directly? If the short name is OS generated, then, it will always contain tilde (assuming that the file name doesn’t fit in 8.3) right?

My concern was actually around backward compatibility. Our filter is consumed by third parties and so far we were returning them normalized names. With the change, we will start returning opened names instead and hence I wanted to understand the differences and see if it can affect them.

We are trying to keep the deviation minimal. We are returning opened names only for network files. If the opened name contains tilde, then, we return normalized name.

Based on the discussion so far, there are few cases which we need to worry about.

  1. An explicit short name was set which doesn’t not contain tilde and the file was originally opened using short name. We will end up sending short filename in this case.
  2. Mount points are not resolved. We will end up returning path containing mount point instead of resolved path.
  3. If the file is opened by id. I looked at Ladislav’s sample to open file by id. However, apparently, it’s not possible to open the file on network with id? Hence, this is probably a mute point?
  4. Any other case that you can think off?

Thanks.
-Prasad