Very slow throughput writing to disk

NTFSD
1

Hello, experts.

I’m a filesystem developer new to Windows. I’ve written a
homegrown metadata cache for my fs. It basically seems to work but
I find that on Windows, when I flush my metadata cache to my log file,
the performance is terrible. I flushed to my log when I got
to ~1MB of dirty metadata. The buffers in my homegrown
cache are 512 bytes each. So I’m flushing about 2000 of
them. The flush takes about 15 seconds or a throughput
of about 70kb/second. Here is my flush algorithm:

For each dirty 512-byte buffer
KeInitializeEvent( &event);
irp = IoBuildAsynchronousFsdRequest( IRP_MJ_WRITE,
deviceObject,
(void *)Buffer,
Length,
&Offset,
NULL );
IoSetCompletionRoutine( irp,…);
IoCallDriver(irp);

This kicks them all off asynchronously, as I understand it.
Then I wait for all the I/O’s to complete. I try to be smart by
waiting on the last one I started:

while (not done)
KeWaitForSingleObject(event)

My I/O completion routine marks my metadata buffers as
being “done” so I only wait for those that are still pending.

As I said, this does seem to work. But the performance is
so bad, I feel like I’m missing a basic Windows concept.
This algorithm works well on other platforms.

Any help would be greatly appreciated!

2

Tim,

  1. Make your buffers 4k each and that will probably give you 8x
    performance.

  2. Better yet, make them 64k and that will probably give you 128x
    performance.

Jerry.

xxxxx@lists.osr.com wrote on 03/13/2006 02:45:30 PM:

Hello, experts.

I’m a filesystem developer new to Windows. I’ve written a
homegrown metadata cache for my fs. It basically seems to work but
I find that on Windows, when I flush my metadata cache to my log file,
the performance is terrible. I flushed to my log when I got
to ~1MB of dirty metadata. The buffers in my homegrown
cache are 512 bytes each. So I’m flushing about 2000 of
them. The flush takes about 15 seconds or a throughput
of about 70kb/second. Here is my flush algorithm:

For each dirty 512-byte buffer
KeInitializeEvent( &event);
irp = IoBuildAsynchronousFsdRequest( IRP_MJ_WRITE,
deviceObject,
(void *)Buffer,
Length,
&Offset,
NULL );
IoSetCompletionRoutine( irp,…);
IoCallDriver(irp);

This kicks them all off asynchronously, as I understand it.
Then I wait for all the I/O’s to complete. I try to be smart by
waiting on the last one I started:

while (not done)
KeWaitForSingleObject(event)

My I/O completion routine marks my metadata buffers as
being “done” so I only wait for those that are still pending.

As I said, this does seem to work. But the performance is
so bad, I feel like I’m missing a basic Windows concept.
This algorithm works well on other platforms.

Any help would be greatly appreciated!

Questions? First check the IFS FAQ at https://www.osronline.
com/article.cfm?id=17

You are currently subscribed to ntfsd as: xxxxx@attotech.com
To unsubscribe send a blank email to xxxxx@lists.osr.com