Very Computer

I recently removed Windows 98 from one of my PCs and replaced it with
Linux (Red Hat 6.2). Before doing that I of course had to copy all
files I wanted to keep from its Windows disk to another; the only such
files were ~100 megabyte .zip files (backups), and I simply used the
destination Win98 box to do the copying (via Windows networking, of
course).

The other PC is called Fred; Lucy is the PC which is changing
Win98=>Linux. The backups (Lucy/Win98->Fred/Win98) copied over my
10BaseT network with an average rate of about 250 kByte/sec, measured
using the System Monitor on Fred and watching File-system bytes
written. While disappointing (theoretical maximum is about 4 times
that rate; the Ethernet is otherwise idle as are both Fred and Lucy),
this is what it is.

I then did a complete install of Linux (Red Hat 6.2) onto Lucy, and
got both networking and Samba working; Fred can mount a user's home
from Lucy. I created a 10 mB file in my home and copied it to Fred
in the same manner as before. This time the average rate is only about
80 kByte/sec. THAT'S TERRIBLE!!! This poor performance makes using
Linux as a local disk server almost useless.

Note that during these copies, Lucy/Win98 (the source) claimed its
CPU is busy about 70%; Lucy/Linux (also the source) claimed about 5%
(or less). For both copies, Fred/Win98 (the destination, doing the
copy) claimed its CPU is 100% busy. On Fred, the disk-writing rate
was steady at 250 kB/sec for Win98->Win98, but alternated each second
between 125 kB/sec and 62 kB/sec for Linux->Win98.

Do other people get better performance than this from Linux? Is this
bottleneck likely to be the network interfaces (Lucy's is an ISA card
using the driver ne.o)? Is this likely to be Samba? Or is it likely to
be my IDE hard drive performance? How can I tune this better?

But most importantly: How can Linux hope to succeed if it gives poorer
performance _AS_A_SERVER_ than Win98??? Remember, this comparison used
exactly the same hardware:

        Fred: Micron PC, Pentium 166 MHz, PCI network card.
        Lucy: Compaq PC, Pentium 233 MHz, ISA network card.
        Both have IDE drives only; Fred's destination is 8.6 GB
        (~50% full), Lucy's source is 4.1 GB (~80% full on Win98,
        ~20% full on Linux); all drives were checked and defragmented
        just before these copies were performed. Under Linux, pings
        Lucy->Fred->Lucy take an average of 0.8 ms, and a "flood
        ping" gets about 1000 packets per second.

What does the kernel say about the ide controller ?
In what mode is the ide controller on your linux machine runing ?

I had a similar problem when I first started working with Samba and Windows
2000.  In the end, I upgraded to the latest version of Samba and it seemed
to fix everything. In fact, it really flys now.  I asked the samba lists
about this and so many variables came back that it was overwhelming.  So, I
would say upgrade to the latest version of Samba and try again.

--
                         Jason
          www.cyborgworkshop.com
...and the geek shall inherit the earth...

  In a message on Thu, 10 Aug 2000 12:05:04 -0500, wrote :

TR> I recently removed Windows 98 from one of my PCs and replaced it with
TR> Linux (Red Hat 6.2). Before doing that I of course had to copy all
TR> files I wanted to keep from its Windows disk to another; the only such
TR> files were ~100 megabyte .zip files (backups), and I simply used the
TR> destination Win98 box to do the copying (via Windows networking, of
TR> course).
TR>
TR> The other PC is called Fred; Lucy is the PC which is changing
TR> Win98=>Linux. The backups (Lucy/Win98->Fred/Win98) copied over my
TR> 10BaseT network with an average rate of about 250 kByte/sec, measured
TR> using the System Monitor on Fred and watching File-system bytes
TR> written. While disappointing (theoretical maximum is about 4 times
TR> that rate; the Ethernet is otherwise idle as are both Fred and Lucy),
TR> this is what it is.
TR>
TR> I then did a complete install of Linux (Red Hat 6.2) onto Lucy, and
TR> got both networking and Samba working; Fred can mount a user's home
TR> from Lucy. I created a 10 mB file in my home and copied it to Fred
TR> in the same manner as before. This time the average rate is only about
TR> 80 kByte/sec. THAT'S TERRIBLE!!! This poor performance makes using
TR> Linux as a local disk server almost useless.
TR>
TR> Note that during these copies, Lucy/Win98 (the source) claimed its
TR> CPU is busy about 70%; Lucy/Linux (also the source) claimed about 5%
TR> (or less). For both copies, Fred/Win98 (the destination, doing the
TR> copy) claimed its CPU is 100% busy. On Fred, the disk-writing rate
TR> was steady at 250 kB/sec for Win98->Win98, but alternated each second
TR> between 125 kB/sec and 62 kB/sec for Linux->Win98.
TR>
TR> Do other people get better performance than this from Linux? Is this
TR> bottleneck likely to be the network interfaces (Lucy's is an ISA card
TR> using the driver ne.o)? Is this likely to be Samba? Or is it likely to
TR> be my IDE hard drive performance? How can I tune this better?
TR>
TR> But most importantly: How can Linux hope to succeed if it gives poorer
TR> performance _AS_A_SERVER_ than Win98??? Remember, this comparison used
TR> exactly the same hardware:
TR>
TR>  Fred: Micron PC, Pentium 166 MHz, PCI network card.
TR>  Lucy: Compaq PC, Pentium 233 MHz, ISA network card.
TR>  Both have IDE drives only; Fred's destination is 8.6 GB
TR>  (~50% full), Lucy's source is 4.1 GB (~80% full on Win98,
TR>  ~20% full on Linux); all drives were checked and defragmented
TR>  just before these copies were performed. Under Linux, pings
TR>  Lucy->Fred->Lucy take an average of 0.8 ms, and a "flood
TR>  ping" gets about 1000 packets per second.

I would guess it is the IDE hardware.  Remember: Linux only has
*minimal* support for IDE/DMA drives -- Win* has better support for
this.

Note: *real* servers NEVER USE IDE DRIVES.  They use SCSI.
Linux w/SCSI has much better performance, esp. with a decent card and a
good disk.

Also, the NE2000 ISA Ethernet cards are really low-end 'junk' cards.
Why do you have such a junky card in a P233 box?  Do you really have ALL
of its PCI slots used up?  At *least* get a 3Com 509B card!  

TR>
TR>

TR>                                

--
                                     \/


http://www.deepsoft.com              /\FidoNet:    1:321/153

Hardly. With proper DMA support compiled into the kernel and (possibly)
some tweaks with hdparm you can get much better performance out of an ATA/
(U)DMA drive under Linux than Win'**.

Quote:>Note: *real* servers NEVER USE IDE DRIVES.  They use SCSI.

That's hardly an answer, and SCSI is a lucury for home users. I'd rather
get 40GB of ATA drive space than 20GB of SCSI drivespace, thank-you.

Quote:>Linux w/SCSI has much better performance, esp. with a decent card and a
>good disk.

Now you're getting to the heart of the matter; quality hardware. If he got
quality ATA drives on a good controller (Ultra ATA66, for example) he could
see massive performance from his system.

My Ultra ATA33 drives could/would send at well over 700kB/Sec over 10Base-T
network cards, from my ISA NE2000 (at the time) to a PCI D-Link.

Quote:>Also, the NE2000 ISA Ethernet cards are really low-end 'junk' cards.
>Why do you have such a junky card in a P233 box?

I'm sure he appreciates the sentiment. In reality, not everybody has the money
to be purchasing $80 NICs for all their machines. It would be nice, but it's
not very realistic.

Quote:>Do you really have ALL of its PCI slots used up?  At *least* get a 3Com 509B
>card!

While you're sending out all this money along with your quality advice, I'll
take an Ultra-2 160 SCSI controller and a 20GB SCSI drive, along with a high
quality NIC. E-Mail me for my mailing address.

--


Humming along under SuSE 6.4, Linux 2.4.0-test5

Quote:>I then did a complete install of Linux (Red Hat 6.2) onto Lucy, and
>got both networking and Samba working; Fred can mount a user's home
>from Lucy. I created a 10 mB file in my home and copied it to Fred
>in the same manner as before. This time the average rate is only about
>80 kByte/sec. THAT'S TERRIBLE!!! This poor performance makes using
>Linux as a local disk server almost useless.

Yes it is.  It should not be that slow.  Since you have IDE drives,
that would be a good place to start.  You'll need to tweak them using
hdparm. The default settings are very conservative.

I use:  hdparm -X66 -d1 -u1 -m16 -c3 /dev/hda

And I get:

/dev/hda:
 Timing buffer-cache reads:   128 MB in  1.68 seconds =76.19 MB/sec
 Timing buffered disk reads:  64 MB in  6.71 seconds = 9.54 MB/sec

This is about double the performance of the default settings.

Quote:>Do other people get better performance than this from Linux? Is this

Yes, much better.

Quote:>bottleneck likely to be the network interfaces (Lucy's is an ISA card
>using the driver ne.o)? Is this likely to be Samba? Or is it likely to
>be my IDE hard drive performance? How can I tune this better?

Ne2000 cards are not good choices for servers.  They use programmed IO,
which uses more cpu time than you'd probably like.  But I'd start with
the IDE tweaks.  See "man hdparm".

Another thing is Samba.  Setting "socket options = TCP_NODELAY" in the
config file can help performance on LAN's.

--
 -| Bob Hauck
 -| Codem Systems, Inc.
 -| http://www.codem.com/