Caching vs non-caching SCSI (was 386/ESDI vs Pentium/SCSI)

Caching vs non-caching SCSI (was 386/ESDI vs Pentium/SCSI)

Post by Lucky Leavel » Sun, 30 Jul 1995 04:00:00



I really don't want to get into a long theoretical discussion.  About two
years ago, I replaced my ADaptec 1542B with an Alpha DC-800 with 4MB on
board.  Performance improvements ranged from roughly 1/3 (copying large
files) to 300% (bringing up X-windows session).  RDBMS (Ingres)
performance improved markedly.  (I do have a 1200 watt UPS, which I would
have with caching or non-caching controllers.)  

I am running OSE 3.0 with 16MB RAM on a 486-50DX2; 3 SCSI HD's, one tape,
one CD-ROM.  I did not attempt to add RAM for a comparison there.

Regards,
Lucky

Lucky Leavell                            Phone: (812) 945-6555
Relational Information Systems, Inc.       FAX: (812) 949-9233

New Albany, IN 47150-2013                       71534,2674 (CompuServe)
                                 WWW Home Page:  http://wl.iglou.com/ris

 
 
 

Caching vs non-caching SCSI (was 386/ESDI vs Pentium/SCSI)

Post by Lucky Leavel » Sun, 06 Aug 1995 04:00:00



> place. A 1542CF should have made an improvement as well, and unless your
> disks are pretty slow, a 2742 or 2940 would probably give a vastly
> improved transfer speed. (The best I've managed for an ISA card under
> IOZONE is about 1.2MB/sec read, while the EISA and PCI cards can easily
> pump 2.5MB/sec+ through per drive, if the drive is up to it.)

The 2742/2940 are out for now, I only have an ISA bus machine (an aging
<G> 496-50DX2).  However, the original question was posed for a client
who was replacing his 386/ESDI with a Pentium-90/SCSI (2940W).  Looks
like he should see at least as much improvement as I did (upwards of 300%).

Now for a question I posted in response to a suggestion that I use 16Mb
(or so) of RAM for I/O buffering.  Would that imply manually setting NBUF
(and related parameters) manually as opposed to allowing SCO to set it at
boot time?

TIA,
Lucky

Lucky Leavell                            Phone: (812) 945-6555
Relational Information Systems, Inc.       FAX: (812) 949-9233

New Albany, IN 47150-2013                       71534,2674 (CompuServe)
                                 WWW Home Page:  http://wl.iglou.com/ris

 
 
 

Caching vs non-caching SCSI (was 386/ESDI vs Pentium/SCSI)

Post by Jeff Lieberma » Tue, 08 Aug 1995 04:00:00


: Now for a question I posted in response to a suggestion that I use 16Mb
: (or so) of RAM for I/O buffering.  Would that imply manually setting NBUF
: (and related parameters) manually as opposed to allowing SCO to set it at
: boot time?

SCO 3.2v4.2 is delivered with a MAXBUF=600 value as the
default.  This means that the boot time NBUF=0 auto configuration
will not grab more than 600KBytes of NBUF disk bufferin.
This is fine for a 4MB ram system and is rediculously small
for todays megabloat servers.

At first, I thought that increasing the value of MAXBUF was the
right way and to let the kernel decide how much it wants.  Not
a good idea.  NHBUF is suppose to be about 1/4 of NBUF and the
auto configuration doesn't change this.  Therefore, set NBUF
to whatever you consider reasonable, set NHBUF to approx 1/4
of this (rounded off to a binary number), and leave MAXBUF alone.

A few things have changed in the last few years.  Processors
have increased in speed by factors of 4 to 10, while disk drives
have improved in speed (data transfer rate) by factors of 2-3.
This tends to create a bottleneck in the drive subsystem.  I
recently upgraded a 486 server to a Pentium 90 and was cursed
with SLOWER performance.  The difference was that the 486 had
16MB of hardware cache on the SCSI adapter, while the Pentium
only had 4MB.  After much seat of the pants testing using Unify
DBM as a benchmark, I ended up with 16MB of SCSI adapter cache,
NBUF=12000, NHBUF=4096, and the remaining 48MB of ram was left
to the OS.  That's 1/3 of the total memory in disk buffers and
it is FAST!  Sar's waiting for i/o (an indication of a disk
bottleneck) went from 80% to 15% with about 80 users pounding
away.

This exercise gave me a chance to test the merits of having all
the disk cache in main ram as NBUF and NHBUF, verses the same
amount on a cacheing controller.  My tests were far from rigorous
but my *GUESS* is that with 64MB of total ram, the cacheing
controller (DPT 2122) is about 15% faster because it can be
readin/rightin the data to/from the disk while the processor
is doing other things.  A crude form of multitasking but effective.

I also had a chance to test various caching adapters.  The
EISA Mylex DAC960E is about 15% faster than the DTC 2122.
However, the Mylex is tended to hang the system with a parity
check when faced with heavy traffic (HP C1533 DAT write to disk)
while the DPT is solid and quite reliable with ECC ram.

There is always the concern about what happens if the power fails.
When the power fails and all the data in BOTH the NBUF and
the hardware cache disappear, you're going to lose your data
and your job.  I think nothing of buying $1000 SPS's to insure
integrity.  That's cheap insurance.  IMHO, there is no way that
you can run todays servers without a monster buffer making
protection manditory.  You should also consider error correcting
(ECC) ram on both the main memory and harware cache.  A one bit
error can ruin your whole day (and night).

--
# Jeff Liebermann   Box 272     1540 Jackson Ave     Ben Lomond    CA   95005

# 408.699.0483 digital_pager    73557,2074  cis [don't]

 
 
 

Caching vs non-caching SCSI (was 386/ESDI vs Pentium/SCSI)

Post by Kevin Smi » Wed, 06 Sep 1995 04:00:00



}...  After much seat of the pants testing using Unify
}DBM as a benchmark, I ended up with 16MB of SCSI adapter cache,
}NBUF=12000, NHBUF=4096, and the remaining 48MB of ram was left
}to the OS.  That's 1/3 of the total memory in disk buffers and
}it is FAST!  Sar's waiting for i/o (an indication of a disk
}bottleneck) went from 80% to 15% with about 80 users pounding
}away.
}...
}There is always the concern about what happens if the power fails.
}When the power fails and all the data in BOTH the NBUF and
}the hardware cache disappear, you're going to lose your data
}and your job.  I think nothing of buying $1000 SPS's to insure
}integrity.  ...

Has anyone experimented with different values of NAUTOUP and
BDFLUSHR and how they effect performance?  This seems particularly
relevant when balancing performance and exposure with a monster
cache.
--

  /\     /\        ___ |[]|_n_n_I_c    ShadeTree Software, Inc.
 /\ '  _`\ `_(==  |___||__|###|____)  Voice: 215-487-3811  Email: !shady!kevin

 
 
 

1. Pentium/SCSI vs 386/ESDI Performance

Be kind to an old fellow but I have a potential sale of a DEC Prioris 590
HX / SCO system configured to replace a 386/33 ISA with single 300Mb ESDI
drive supporting (very slowly) about 16 RealWorld accounting software
users under AT&T Unix.  The DEC system has 48MB memory, two 1-Gb drives
attached to the onboard PCI SCSI controller and uses the Digiboard XEM
multiport board.  The AT&T system has 24Mb memory.

   1. For the amount of money involved, they want some kind of assurances
        the new system will solve their performance problems; I am reluctant
        to make specific claims because 1) I don't know and 2) I don't want
        to misrepresent the new system.

        NOTE: sar stats on the old system show the ESDI drive busy an
        average of 16% with average I/O wait time of 150-200ms with peaks over
        500.  Run queue occupancy averaged 10-12 100% of the time. There was
        no swapping and paging was within acceptable limits. The average
        CPU wait time is 77% which rarely drops below 50%.

        NOTE: A different customer recently upgraded from a 386/25 to a P90
        on a much lighter loaded system which ran RealWorld and a RDBMS and
        TCP/IP.  Their sar stats look very good but, unfortunately, we
        don't have any on the old system to compare with (unless they still
        have that master backup I made just before the old system died).

    2.  I would love to have a caching SCSI controller, but, unfortunately,
        DEC doesn't offer one; the system must be single sourced to meet the
        vendor-supplied service requirement (it will be located 2500 miles
        away).

    3.  If you don't have experience with the DEC Prioris HX, perhaps similar
        experiences would suffice.

If you have something to share and wouldn't mind a phone call this
afternoon from the prospective customer, please email me your name and
number.

Thank you,
Lucky

Lucky Leavell                            Phone: (812) 945-6555
Relational Information Systems, Inc.       FAX: (812) 949-9233

New Albany, IN 47150-2013                       71534,2674 (CompuServe)
                                 WWW Home Page:  http://wl.iglou.com/ris

2. Help needed on storing file on disk from Unix???

3. 256k cache vs 512k cache

4. Problem with Promise Ultra33

5. ???---RAM vs cache : is 256k cache enough for 24M RAM ---???

6. Adaptec 2100S RAID controller on FreeBSD

7. Buffer Cache vs. Page Cache?

8. Bug in Berkley db code?

9. 512K cache vs. 256K cache

10. EDO vs FP ram, async vs burst-mode cache

11. Linux vs OS2 vs NT vs Win95 vs Multics vs PDP11 vs BSD geeks

12. FPU performance of Pentium vs. K6 vs. Ppro vs. PII

13. Trident vs ET4000 for X11 (386 vs 486)