Awful FS performance?

Am I doing something wrong, or does RtP have really bad
filesytem performance? On Bonnie, it benchmarks at
almost a third as high as Linux’s ResiesFS, and well
under half of BeOS’s bfs. In iozone, ReiserFS gives me
28MB/sec reads, while QNX gives me 5-6MB/sec reads. I
have UltraDMA turned on, and it IS called QNX FFS (Fast
File System) so whats up? I heard that by default BSD-
type filesystems do synchronus writes, but that wouldn’t
affect iozone (a sequential) benchmark, would it? And
BSD (not running softupdates) gets a full 19MB/sec
faster than QNX on iozone (24 vs 5)


Sent via Deja.com http://www.deja.com/
Before you buy.

Previously, heliosc@my-deja.com wrote in comp.os.qnx:

Am I doing something wrong, or does RtP have really bad
filesytem performance?

Filesystem performance is fine, personally, the benchmarks are useless
unless they represent the normal activity on your system. If they don’t
then they are just useless numbers… anyhow, with that aside…

have UltraDMA turned on, and it IS called QNX FFS (Fast
File System) so whats up?

QNX does not support the UltraDMA modes of ATAPI/IDE, so that is most
likely the reason for the dramatic change is performance. As for FFS,
you might want to check that… last I heard that was for “Flash File
System” and is typically used for writing to flash based memory cards
such as a PCMCIA flash card.

Cheers,
Camz.

–
Martin Zimmerman camz@passageway.com
Camz Software Enterprises www.passageway.com/camz/qnx/
QNX Programming & Consulting

Filesystem performance is fine, personally, the benchmarks are useless
unless they represent the normal activity on your system. If they don’t
then they are just useless numbers… anyhow, with that aside…

Benchmarks are not useless. I do a lot of things with the system, and
transferring large files is one of them. Given the fact that Photon seems so
adept at handling graphics, it would seem that fast handling of video data
would be an ideal use for QNX. However, if the FS can’t live up to that…


QNX does not support the UltraDMA modes of ATAPI/IDE, so that is most
likely the reason for the dramatic change is performance. As for FFS,
you might want to check that… last I heard that was for “Flash File
System” and is typically used for writing to flash based memory cards
such as a PCMCIA flash card.

Cheers,
Camz.

Really? Why not? Several places in the documentation it refers to UltraDMA,
and it allows me to change the DMA settings at bootup. I’m talking about QNX
RtP here. As for FFS, it might stand for flash file system, but in the large
majority of cases, it stands for Berkley Fast File System, the FS derived
from the original UFS and used on many commercial UNIXes today.


Sent via Deja.com http://www.deja.com/
Before you buy.

The QNX RtP doesn’t use FFS, it uses PFS, which is the package file system.


<heliosc@aircor.net> wrote in message news:90jph3$sdp$1@nnrp1.deja.com…

Filesystem performance is fine, personally, the benchmarks are useless
unless they represent the normal activity on your system. If they don’t
then they are just useless numbers… anyhow, with that aside…

Benchmarks are not useless. I do a lot of things with the system, and
transferring large files is one of them. Given the fact that Photon seems
so
adept at handling graphics, it would seem that fast handling of video data
would be an ideal use for QNX. However, if the FS can’t live up to that…


QNX does not support the UltraDMA modes of ATAPI/IDE, so that is most
likely the reason for the dramatic change is performance. As for FFS,
you might want to check that… last I heard that was for “Flash File
System” and is typically used for writing to flash based memory cards
such as a PCMCIA flash card.

Cheers,
Camz.

Really? Why not? Several places in the documentation it refers to
UltraDMA,
and it allows me to change the DMA settings at bootup. I’m talking about
QNX
RtP here. As for FFS, it might stand for flash file system, but in the
large
majority of cases, it stands for Berkley Fast File System, the FS derived
from the original UFS and used on many commercial UNIXes today.


Sent via Deja.com > http://www.deja.com/
Before you buy.

PFS is something else entirely. I’m talking about the main filesystem (which
stores files and apps as opposed to PFS which is a driver that simulates a
package filesystem, and writes to the underlying QNX FS. Anyway, it seems
that I’m mistaken, I had heard somewhere that it used a version of Berkley’s
FFS (like BSD), but I can’t find anything to back that up. Either way, the FS
is apparently similar to FFS, and given FFS’s awesome performance, QNX
shouldn’t be any different.


The QNX RtP doesn’t use FFS, it uses PFS, which is the package file system.

heliosc@aircor.net> > wrote in message news:90jph3$sdp$> 1@nnrp1.deja.com> …

Filesystem performance is fine, personally, the benchmarks are useless
unless they represent the normal activity on your system. If they don’t
then they are just useless numbers… anyhow, with that aside…

Benchmarks are not useless. I do a lot of things with the system, and
transferring large files is one of them. Given the fact that Photon seems
so
adept at handling graphics, it would seem that fast handling of video data
would be an ideal use for QNX. However, if the FS can’t live up to that…


QNX does not support the UltraDMA modes of ATAPI/IDE, so that is most
likely the reason for the dramatic change is performance. As for FFS,
you might want to check that… last I heard that was for “Flash File
System” and is typically used for writing to flash based memory cards
such as a PCMCIA flash card.

Cheers,
Camz.

Really? Why not? Several places in the documentation it refers to
UltraDMA,
and it allows me to change the DMA settings at bootup. I’m talking about
QNX
RtP here. As for FFS, it might stand for flash file system, but in the
large
majority of cases, it stands for Berkley Fast File System, the FS derived
from the original UFS and used on many commercial UNIXes today.


Sent via Deja.com > http://www.deja.com/
Before you buy.

Sent via Deja.com http://www.deja.com/
Before you buy.

The QNX file system has traditionally emphasized reliability over speed,
so I would not expect QNX to be the fastest filesystem on the block. I
do, however, expect it to be among the most reliable filesystems available
though.

I personally, have no problem with it’s performance (i.e. it is fast enough
for me).

<heliosc@my-deja.com> wrote in message news:90k4n4$5rt$1@nnrp1.deja.com…

PFS is something else entirely. I’m talking about the main filesystem
(which
stores files and apps as opposed to PFS which is a driver that simulates a
package filesystem, and writes to the underlying QNX FS. Anyway, it seems
that I’m mistaken, I had heard somewhere that it used a version of
Berkley’s
FFS (like BSD), but I can’t find anything to back that up. Either way, the
FS
is apparently similar to FFS, and given FFS’s awesome performance, QNX
shouldn’t be any different.


The QNX RtP doesn’t use FFS, it uses PFS, which is the package file
system.

heliosc@aircor.net> > wrote in message
news:90jph3$sdp$> 1@nnrp1.deja.com> …

Filesystem performance is fine, personally, the benchmarks are
useless
unless they represent the normal activity on your system. If they
don’t
then they are just useless numbers… anyhow, with that aside…

Benchmarks are not useless. I do a lot of things with the system, and
transferring large files is one of them. Given the fact that Photon
seems
so
adept at handling graphics, it would seem that fast handling of video
data
would be an ideal use for QNX. However, if the FS can’t live up to
that…


QNX does not support the UltraDMA modes of ATAPI/IDE, so that is
most
likely the reason for the dramatic change is performance. As for
FFS,
you might want to check that… last I heard that was for “Flash
File
System” and is typically used for writing to flash based memory
cards
such as a PCMCIA flash card.

Cheers,
Camz.

Really? Why not? Several places in the documentation it refers to
UltraDMA,
and it allows me to change the DMA settings at bootup. I’m talking
about
QNX
RtP here. As for FFS, it might stand for flash file system, but in the
large
majority of cases, it stands for Berkley Fast File System, the FS
derived
from the original UFS and used on many commercial UNIXes today.


Sent via Deja.com > http://www.deja.com/
Before you buy.




Sent via Deja.com > http://www.deja.com/
Before you buy.

That is not really convincing. The filesystem (which is obsolete too) aside,
RTP still has problem with raw block I/O performance and that has nothing to
do with reliability. Try timing ‘dd if=/dev/hd0 of=/dev/null count=100000’.
On a modern hard EIDE drive in UDMA mode on Intel chipset I get only
4Mb/sec.

The problem lies with drivers. RTP EIDE driver seems to actually work better
on SiS chipset for some weird reason (about twice faster). Now situation
with SCSI is even worse. The aha8 driver is good at reading (about 20Mb/sec
on a good drive) but that’s the only good news. Writing is only about
3Mb/sec. The NCR/Symbios driver gives miserable 2.8Mb/sec either reading or
writing (I guess because it does not work in either sync or wide mode).

Win2k on the same machine on comparable test gives 33Mb/sec reading and
14Mb/sec writing on EIDE and some 14Mb/sec reading/writing on SCSI
(AHA2940UW).

That situation sucks and they gotta do something. Unfortunately their
traditional common language is “we’re not in a mass storage business”.

  • Igor

“John Doe” <john@csical.com> wrote in message
news:90k9dt$8td$1@inn.qnx.com…

The QNX file system has traditionally emphasized reliability over speed,
so I would not expect QNX to be the fastest filesystem on the block. I
do, however, expect it to be among the most reliable filesystems available
though.

I personally, have no problem with it’s performance (i.e. it is fast
enough
for me).

heliosc@my-deja.com> > wrote in message news:90k4n4$5rt$> 1@nnrp1.deja.com> …
PFS is something else entirely. I’m talking about the main filesystem
(which
stores files and apps as opposed to PFS which is a driver that simulates
a
package filesystem, and writes to the underlying QNX FS. Anyway, it
seems
that I’m mistaken, I had heard somewhere that it used a version of
Berkley’s
FFS (like BSD), but I can’t find anything to back that up. Either way,
the
FS
is apparently similar to FFS, and given FFS’s awesome performance, QNX
shouldn’t be any different.


The QNX RtP doesn’t use FFS, it uses PFS, which is the package file
system.

heliosc@aircor.net> > wrote in message
news:90jph3$sdp$> 1@nnrp1.deja.com> …

Filesystem performance is fine, personally, the benchmarks are
useless
unless they represent the normal activity on your system. If they
don’t
then they are just useless numbers… anyhow, with that aside…

Benchmarks are not useless. I do a lot of things with the system,
and
transferring large files is one of them. Given the fact that Photon
seems
so
adept at handling graphics, it would seem that fast handling of
video
data
would be an ideal use for QNX. However, if the FS can’t live up to
that…


QNX does not support the UltraDMA modes of ATAPI/IDE, so that is
most
likely the reason for the dramatic change is performance. As for
FFS,
you might want to check that… last I heard that was for “Flash
File
System” and is typically used for writing to flash based memory
cards
such as a PCMCIA flash card.

Cheers,
Camz.

Really? Why not? Several places in the documentation it refers to
UltraDMA,
and it allows me to change the DMA settings at bootup. I’m talking
about
QNX
RtP here. As for FFS, it might stand for flash file system, but in
the
large
majority of cases, it stands for Berkley Fast File System, the FS
derived
from the original UFS and used on many commercial UNIXes today.


Sent via Deja.com > http://www.deja.com/
Before you buy.




Sent via Deja.com > http://www.deja.com/
Before you buy.

Igor Kovalenko wrote:

That is not really convincing. The filesystem (which is obsolete too) aside,
RTP still has problem with raw block I/O performance and that has nothing to
do with reliability. Try timing ‘dd if=/dev/hd0 of=/dev/null count=100000’.

I’m not sure about qrtp, but on some systems specifying the blocksize
also makes a difference (i.e. “bs=” paramater).

On a modern hard EIDE drive in UDMA mode on Intel chipset I get only
4Mb/sec.

The problem lies with drivers. RTP EIDE driver seems to actually work better
on SiS chipset for some weird reason (about twice faster). Now situation
with SCSI is even worse. The aha8 driver is good at reading (about 20Mb/sec
on a good drive) but that’s the only good news. Writing is only about
3Mb/sec. The NCR/Symbios driver gives miserable 2.8Mb/sec either reading or
writing (I guess because it does not work in either sync or wide mode).

Win2k on the same machine on comparable test gives 33Mb/sec reading and
14Mb/sec writing on EIDE and some 14Mb/sec reading/writing on SCSI
(AHA2940UW).

That situation sucks and they gotta do something. Unfortunately their
traditional common language is “we’re not in a mass storage business”.

  • Igor

“John Doe” <> john@csical.com> > wrote in message
news:90k9dt$8td$> 1@inn.qnx.com> …
The QNX file system has traditionally emphasized reliability over speed,
so I would not expect QNX to be the fastest filesystem on the block. I
do, however, expect it to be among the most reliable filesystems available
though.

I personally, have no problem with it’s performance (i.e. it is fast
enough
for me).

heliosc@my-deja.com> > wrote in message news:90k4n4$5rt$> 1@nnrp1.deja.com> …
PFS is something else entirely. I’m talking about the main filesystem
(which
stores files and apps as opposed to PFS which is a driver that simulates
a
package filesystem, and writes to the underlying QNX FS. Anyway, it
seems
that I’m mistaken, I had heard somewhere that it used a version of
Berkley’s
FFS (like BSD), but I can’t find anything to back that up. Either way,
the
FS
is apparently similar to FFS, and given FFS’s awesome performance, QNX
shouldn’t be any different.


The QNX RtP doesn’t use FFS, it uses PFS, which is the package file
system.

heliosc@aircor.net> > wrote in message
news:90jph3$sdp$> 1@nnrp1.deja.com> …

Filesystem performance is fine, personally, the benchmarks are
useless
unless they represent the normal activity on your system. If they
don’t
then they are just useless numbers… anyhow, with that aside…

Benchmarks are not useless. I do a lot of things with the system,
and
transferring large files is one of them. Given the fact that Photon
seems
so
adept at handling graphics, it would seem that fast handling of
video
data
would be an ideal use for QNX. However, if the FS can’t live up to
that…


QNX does not support the UltraDMA modes of ATAPI/IDE, so that is
most
likely the reason for the dramatic change is performance. As for
FFS,
you might want to check that… last I heard that was for “Flash
File
System” and is typically used for writing to flash based memory
cards
such as a PCMCIA flash card.

Cheers,
Camz.

Really? Why not? Several places in the documentation it refers to
UltraDMA,
and it allows me to change the DMA settings at bootup. I’m talking
about
QNX
RtP here. As for FFS, it might stand for flash file system, but in
the
large
majority of cases, it stands for Berkley Fast File System, the FS
derived
from the original UFS and used on many commercial UNIXes today.


Sent via Deja.com > http://www.deja.com/
Before you buy.




Sent via Deja.com > http://www.deja.com/
Before you buy.

Reading all these comments I can see that there probably are
some problems in the I/O speed to be dealt with. However I
read somewhere here a speed of 28Mb/sec from some benchmark,
and I have to comment.

It’s really simple math. Figure out how fast your drive
spins. Find out the density of sectors per track. Multiply
carefully and you will find that there are no drives that
perform this fast. Not even close. The last time I looked,
Seagate had a 10K rpm drive that would perform in the
10-12Meg/sec area.

RAIDs are an exception. With striped drives, RAID-0 you can
get close to the bus bandwidth.


Mitchell Schoenbrun --------- maschoen@pobox.com

I don’t have anything to do with the file system, and I don’t know what
the issues are that would cause these numbers, but I will put out
that when I extract a large tar file under any linux distro, it’s done fairly
quick, but my system is un responsive for a few seconds after wards (I’m guessing the
driver is flushing it’s cache…) under qnx it takes longer, but I can at
least do stuff.

drempel


heliosc@my-deja.com wrote:

Am I doing something wrong, or does RtP have really bad
filesytem performance? On Bonnie, it benchmarks at
almost a third as high as Linux’s ResiesFS, and well
under half of BeOS’s bfs. In iozone, ReiserFS gives me
28MB/sec reads, while QNX gives me 5-6MB/sec reads. I
have UltraDMA turned on, and it IS called QNX FFS (Fast
File System) so whats up? I heard that by default BSD-
type filesystems do synchronus writes, but that wouldn’t
affect iozone (a sequential) benchmark, would it? And
BSD (not running softupdates) gets a full 19MB/sec
faster than QNX on iozone (24 vs 5)



Sent via Deja.com > http://www.deja.com/
Before you buy.

Andrew Thomas wrote:

…
slower under certain kinds of operations. I have seen many people post
their
qualitative observations, and I have seen many explanations, or at least
suggestions,
that have never worked their way back to a methodical investigation of this
speed
difference. I certainly haven’t put the time into it.

Well, I did. I have tried to do relative measurements on different
systems using Windows and RTP under similar and different tests to find
as much correlation as I can. There certainly are problems and they are
in different areas but all contribute to the result.

  1. Raw drivers performance.

For EIDE the main hurdle is lack of support for UDMA66/100. It also
appears that when controller/drive autonegotiate UDMA66 mode the RTP
driver does not really provide even UDMA33 performance. I get same
results from old ATA drive and new UDMA100 drive on UDMA66 controller.
However on UDMA33 controller the UDMA100 drive works much faster (up to
8Mb/sec read/write). Windows is still much faster anyway (up to 33Mb/sec
reading and 14Mb/sec writing in UDMA66 mode, including filesystem
overhead). That speed should be approacheable even for UDMA33 mode so
there is certainly some other problem in RTP EIDE driver, aside for
UDMA66/100 mode.

For SCSI the trouble is universally bad writing performance and bad
reading performance with some drivers. Some drivers like ncr8 simply do
not support sync/wide mode which is unacceptable. I’m not sure about
btmm driver, docs don’t say a word and I have no hardware to try. With
aha8 the reading speed of 20Mb/sec (better than Win2000 12Mb/sec) is in
such a deep contrast with writing speed of 3Mb/sec that there must be
something wrong. May be it has to do with next issue.

  1. Write caching.

It appears that Windows disk (as opposed to controller) drivers
certainly do something with write caching, which helps a lot. The same
test performed with write cache disabled (on per-drive basis) gives only
marginally better results than on RTP. For example, 5Mb/sec for EIDE and
6Mb/sec for SCSI. With write cache enabled that becomes 14Mb/sec and
11Mb/sec respectively. On RTP (4Mb/sec EIDE, 3Mb/sec SCSI) nothing
helps, not even commit=none mode. Also, judging by the difference in
boost derived from write caching, it appears that caches are different
between EIDE and SCSI, which would mean they have to do with on-drive
buffer (bigger on EIDE) rather than FS cache (which should be
universally effective).

I’m not familiar enough with that stuff, can someone tell me if drivers
can/do indeed control on-drive buffers under Windows? I guess nothing
like that is being done on RTP and that would explain awful write
performance. Or is it just difference between EIDE and SCSI subsystems?

  1. FS performance

It certainly is not optimized for large volumes and high data rates. It
is simply too old and has not been revisited much in last decade. May be
it is ok for embedded market, but sure not for servers/workstations.

Worst of all, there are no even beta/draft specs for disk subsystem
drivers. Nor source available yet, so we can’t figure what is wrong and
can’t write drivers either. They plan on revorking the whole stuff,
adding CAM support and such, may be that’s the chance to get them do it
right.

  • Igor

Hmmm, I wasn’t really trying to “convince” anyone of anything.

I do agree, that from QSSL’s current business perspective, there
is not a particularly hard-driving business case for a HP filesystem,
since this is typically required for file servers, which Linux/NT
address quite well (and who want’s to get involved in that fight :slight_smile:

In the CE space, there is a business case for a reliable and compact
file system, where fast enough is all that matters…

“Igor Kovalenko” <kovalenko@home.com> wrote in message
news:90kmns$fn9$1@inn.qnx.com…

That is not really convincing. The filesystem (which is obsolete too)
aside,
RTP still has problem with raw block I/O performance and that has nothing
to
do with reliability. Try timing ‘dd if=/dev/hd0 of=/dev/null
count=100000’.
On a modern hard EIDE drive in UDMA mode on Intel chipset I get only
4Mb/sec.

The problem lies with drivers. RTP EIDE driver seems to actually work
better
on SiS chipset for some weird reason (about twice faster). Now situation
with SCSI is even worse. The aha8 driver is good at reading (about
20Mb/sec
on a good drive) but that’s the only good news. Writing is only about
3Mb/sec. The NCR/Symbios driver gives miserable 2.8Mb/sec either reading
or
writing (I guess because it does not work in either sync or wide mode).

Win2k on the same machine on comparable test gives 33Mb/sec reading and
14Mb/sec writing on EIDE and some 14Mb/sec reading/writing on SCSI
(AHA2940UW).

That situation sucks and they gotta do something. Unfortunately their
traditional common language is “we’re not in a mass storage business”.

  • Igor

“John Doe” <> john@csical.com> > wrote in message
news:90k9dt$8td$> 1@inn.qnx.com> …
The QNX file system has traditionally emphasized reliability over speed,
so I would not expect QNX to be the fastest filesystem on the block. I
do, however, expect it to be among the most reliable filesystems
available
though.

I personally, have no problem with it’s performance (i.e. it is fast
enough
for me).

heliosc@my-deja.com> > wrote in message
news:90k4n4$5rt$> 1@nnrp1.deja.com> …
PFS is something else entirely. I’m talking about the main filesystem
(which
stores files and apps as opposed to PFS which is a driver that
simulates
a
package filesystem, and writes to the underlying QNX FS. Anyway, it
seems
that I’m mistaken, I had heard somewhere that it used a version of
Berkley’s
FFS (like BSD), but I can’t find anything to back that up. Either way,
the
FS
is apparently similar to FFS, and given FFS’s awesome performance, QNX
shouldn’t be any different.


The QNX RtP doesn’t use FFS, it uses PFS, which is the package file
system.

heliosc@aircor.net> > wrote in message
news:90jph3$sdp$> 1@nnrp1.deja.com> …

Filesystem performance is fine, personally, the benchmarks are
useless
unless they represent the normal activity on your system. If
they
don’t
then they are just useless numbers… anyhow, with that aside…

Benchmarks are not useless. I do a lot of things with the system,
and
transferring large files is one of them. Given the fact that
Photon
seems
so
adept at handling graphics, it would seem that fast handling of
video
data
would be an ideal use for QNX. However, if the FS can’t live up to
that…


QNX does not support the UltraDMA modes of ATAPI/IDE, so that is
most
likely the reason for the dramatic change is performance. As
for
FFS,
you might want to check that… last I heard that was for “Flash
File
System” and is typically used for writing to flash based memory
cards
such as a PCMCIA flash card.

Cheers,
Camz.

Really? Why not? Several places in the documentation it refers to
UltraDMA,
and it allows me to change the DMA settings at bootup. I’m talking
about
QNX
RtP here. As for FFS, it might stand for flash file system, but in
the
large
majority of cases, it stands for Berkley Fast File System, the FS
derived
from the original UFS and used on many commercial UNIXes today.


Sent via Deja.com > http://www.deja.com/
Before you buy.




Sent via Deja.com > http://www.deja.com/
Before you buy.
\

Mario Charest wrote in message …

Same here, but -d is set to 30 seconds. Only draw back to this is when
the cache is flushed, because the data in the cache doesn’t have the
priority
of the client associated with it, the flushing is done at Fsys.eide
priority.

I’ve noticed that something in Fsys re. priorities has changed in the past 3
years. It used to be that the system was still usable when the write-behind
cache was flushing, but now it happens at a much higher priority and user
performance “glitches”. At least, that is my impression.

Oh well, not my concern any more.

That makes the machine slightly irresponsive since Fsys.eide doesn’t
use DMA and is very CPU intensive. But it’s well worth the increase
in performacen. Not to mention a great undelete feature > :wink:

:slight_smile:

–
Bill Flowers
Clearwater, FL (where it is currently 61F compared to Ottawa at 7F or -14C!)

It’s ab interesting coincidence that this discussion is taking place.
I have a local customer who had the following obervations:

============= start of quote ========================
We have tried the various suggestions, but we have not been able to
increase
write performance beyond the 4MB/s mark.

I haven’t managed to test blockwrite/blockread but when I did before, I
got
similar performance. Also iozone gave similar performance.

Parameter changes to Fsys did not make any noticeable diffence,
including
increasing cache size, including write behind delay time and even the
asynchrynous Fsys -a mode

Considering we got 6.2 MB/s under NT (writing to FAT16) and 7+ MB/s
under
dos, I find it hard to believe that we can only achieve 4MB/s under QNX
(only 3.1 MB/s if we write to FAT16 instead of the QNX file system).


============== end of quote =============================

their performance benchmarks showed the following:

I have tested using a faster system (P2 450MHz - our SBC is running a
Pentium (1) 233MHz) and I have got the following performance figures
writing
to a FAT16 partition

QNX 6 MB/s
NT 7.8 MB/s
DOS 8.5 MB/s

compared to the 233 MHz Pentium

QNX 3.3 MB/s
NT 6.2 MB/s
DOS 7.9 MB/s

For several years the QNX website mentioned “platter speed” performance
wrt to the Fsys drivers, but this patently is not so (and the wording
has now been removed). I’m not sure that it ever was the case?
Possibly what happened is that years ago the IDE was IO bound,
and with slowish CPU’s of the time, the CPU was idle waiting
for the disk.
DMA was never implemented (why?), and no one noticed.
Nowadays, the IDE is much faster, and the CPUs can’t keep
up (without DMA anyway). Now it is very obvious. A 400MHz CPU
will basically get double the Fsys performance compared to
a 200 MHz processor under QNX4, proving QNX4 Fsys is now CPU bound.

The observation about the Fsys performance being what is (was?)
good enough for embedded is accurate I think. No one noticed
until my customer did that the QNX4 Fsys performance now sucks compared
to NT, DOS and RTP.

Whether this is still going to be the case is an open question.
My customer’s application is requiring the max possible throughput.
Their options now are to use the fastest possible CPU, or port to RTP.

Bill-To satisfy my idle curiosity, why was DMA never implemented under
QNX4? Realtime considerations? If so, the RTP option of being able to
disable it seems more useful…

–
Systems 104 (South Africa) http://www.systems104.co.za


Sent via Deja.com http://www.deja.com/
Before you buy.

IMHO it makes sense in embedded system where a HD is needed.
If a Pentim 200 is required to hit 5M/sec and a more performance
oriented FS/driver would allow Pentium 100Mzh to hit that same
5M, it’s a real plus.

But if QSSL made the decision that HD speed is not important
for their market, it’s up to them.

Also, some embedded apps need to move large amounts of data around, for
example logging type applications. I read a post from a developer (either on
this board or qnxstart) who was using QNX to write a logging app, and the
data rate of the logging overwhelemed the filesystem performance, even though
the HD was more than fast enough. Also, I see this as missed potential. Many
good filesystems exist, and getting a top notch one should involve little
more than “borrowing” BSD’s FFS +SoftUpdates system and adapting it to RtP.
Yea it would be a bit of work, but for the payoff I think it would be useful.
(Yea, I’m probably underestimating the work of porting FFS, but if they can
make FAT16 work with an RTOS, FFS should be a piece of cake.)


Sent via Deja.com http://www.deja.com/
Before you buy.

“W.A. (Bill) Flowers” <wflowers_NOSPAM@gte.net> wrote in message
news:W4NX5.639$h06.80133@dfiatx1-snr1.gtei.net…

Are we talking about RTP (aka Neutrino) or QNX4? I can’t comment on what
is
happening in RTP because I left QSSL while RTP was still in its infancy.
I
haven’t a clue as to how the RTP Fsys is tuned by default or how to change
its tuning.

I’m talking about RTP and it has much the same options for tuning. Using
async mode does not help a bit. Did not try increasing flush delay yet.

Under QNX4 the Fsys is tuned for reliability by default. However for my
desktop I never run the defaults, instead I tune for performance (and use
a
UPS for security). In particular I increase the write-behind delay time
(‘-d …’) to at least 5 seconds and turn on the ‘-a’ (asynchronous
metadata
update) option. Since 4.25, the ‘-a’ option can be invoked as a mount
option.

Could not find a way to pass it to mount in RTP.

  • igor

Try different drives. And try options, especially those related to drive
geometry, I just guess that’s where it could do any math. Since it does not
ever detect geometry properly, I always pass it this:

‘disk translation=heads:cyls:path:target:lun’ (not sure about order, see
docs).

  • igor

“Eric Doenges” <Eric.Doenges@lpr.ei.tum.de> wrote in message
news:doenges.976086406@lpr.e-technik.tu-muenchen.de…

“Igor Kovalenko” <> kovalenko@home.com> > writes:

[ … ]
The problem lies with drivers. RTP EIDE driver seems to actually work
better
on SiS chipset for some weird reason (about twice faster). Now situation
with SCSI is even worse. The aha8 driver is good at reading (about
20Mb/sec
on a good drive) but that’s the only good news. Writing is only about
3Mb/sec. The NCR/Symbios driver gives miserable 2.8Mb/sec either reading
or
writing (I guess because it does not work in either sync or wide mode).

That’s still a lot better than what I’m getting – on my machine, the
NCS/symbios driver terminates with a floating point (!) exception, which
means I can’t access my CD-ROM, which in turn means QNX RtP is currently
totally useless for me.


Dipl.-Ing. Eric Doenges > http://www.rcs.ei.tum.de
Institute for Real-Time Computer Systems (RCS) fon +49-89-289-23590
Technische Universitaet Muenchen, D-80290 Muenchen fax +49-89-289-23555

<heliosc@my-deja.com> wrote in message news:90p2gn$6f8$1@nnrp1.deja.com…

IMHO it makes sense in embedded system where a HD is needed.
If a Pentim 200 is required to hit 5M/sec and a more performance
oriented FS/driver would allow Pentium 100Mzh to hit that same
5M, it’s a real plus.

But if QSSL made the decision that HD speed is not important
for their market, it’s up to them.

Also, some embedded apps need to move large amounts of data around, for
example logging type applications. I read a post from a developer (either
on
this board or qnxstart) who was using QNX to write a logging app, and the
data rate of the logging overwhelemed the filesystem performance, even
though
the HD was more than fast enough.

Also, I see this as missed potential.

I totaly agree with you Igor, however I also see

  • no wireless (I know Chris…)
  • no firewire
  • no 1GigaBit Ethernet
  • no 3D aside 3Dfx
  • no USB (comming soon though)
  • no decent email program
  • slow compiler
  • no integrate environment
  • no kernel instrumentation
  • no support for RAID HD controler (could be a easy way to increase
    performance)
  • no HTML 4.0

as missed opportunity… Given the number of employe they currently
have, it’s amazing what they have come up with already. They can
keep everyone of us 100% happy with every aspect of the product.
However I’m all for keeping on pushing them to improve, as long
as it’s positive and constructive.



Many

good filesystems exist, and getting a top notch one should involve little
more than “borrowing” BSD’s FFS +SoftUpdates system and adapting it to
RtP.
Yea it would be a bit of work, but for the payoff I think it would be
useful.
(Yea, I’m probably underestimating the work of porting FFS, but if they
can
make FAT16 work with an RTOS, FFS should be a piece of cake.)


Sent via Deja.com > http://www.deja.com/
Before you buy.

True. However there are primary things and secondary. People will not bother
even to try most of what you’ve mentioned if they know that such a basic
thing is so poorly implemented. A PC starts with video, disk, network,
audio. If that works satisfactory, people may think about playing with the
rest. RTP offered major improvements in all those areas except disk
subsystem. What is bothering me, I don’t see Kevin talking here or anyone
else who might be involved in development.

  • Igor

“Mario Charest” <mcharest@void_zinformatic.com> wrote in message
news:Dj_X5.1534$He.88420@wagner.videotron.net…

I totaly agree with you Igor, however I also see

  • no wireless (I know Chris…)
  • no firewire
  • no 1GigaBit Ethernet
  • no 3D aside 3Dfx
  • no USB (comming soon though)
  • no decent email program
  • slow compiler
  • no integrate environment
  • no kernel instrumentation
  • no support for RAID HD controler (could be a easy way to increase
    performance)
  • no HTML 4.0

as missed opportunity… Given the number of employe they currently
have, it’s amazing what they have come up with already. They can
keep everyone of us 100% happy with every aspect of the product.
However I’m all for keeping on pushing them to improve, as long
as it’s positive and constructive.



Many
good filesystems exist, and getting a top notch one should involve
little
more than “borrowing” BSD’s FFS +SoftUpdates system and adapting it to
RtP.
Yea it would be a bit of work, but for the payoff I think it would be
useful.
(Yea, I’m probably underestimating the work of porting FFS, but if they
can
make FAT16 work with an RTOS, FFS should be a piece of cake.)


Sent via Deja.com > http://www.deja.com/
Before you buy.

“Igor Kovalenko” <kovalenko@home.com> writes:

Try different drives. And try options, especially those related to drive
geometry, I just guess that’s where it could do any math. Since it does not
ever detect geometry properly, I always pass it this:

‘disk translation=heads:cyls:path:target:lun’ (not sure about order, see
docs).

I thought drive geometry was handled in the SCSI device itself ? Or is QNX
RtP one more example of SCSI just not working correctly on PCs ?


Dipl.-Ing. Eric Doenges http://www.rcs.ei.tum.de
Institute for Real-Time Computer Systems (RCS) fon +49-89-289-23590
Technische Universitaet Muenchen, D-80290 Muenchen fax +49-89-289-23555