Conference ssag::ask_ssag

6376.0. "KZPSC Bandwidth again" by JOBURG::FOLLY () Thu Feb 06 1997 05:49

    	Hi,
    
    I am entering a new note on KZPSC bandwidth although there has been a
    discussion in 4633. But the replies are old and contradict the
    statement in Mylex homepage. I need to know the bandwidth provided by
    KZPSC. The product description on www.storage.digital.com does not give
    any info. In short the reply to 4633 says the aggregate bandwidth is 
    18 MB/s and Mylex says it is 60 MB/s. 
    Can someone clarify who's right ?
    
    Thanks
    
    Hereafter is what Mylex says on his homepage
    
EXTRACT FROM MYLEX Homepage

http://www.mylex.com/products/prodinfo/pd1.html

................................
Premier Performance 

     The DAC960PD controller lets you take full advantage of 132 MB/sec PCI 
technology to increase I/O
     performance for network server and storage warehousing applications. An 
Intel i960� 32-bit RISC
     processor drives all disk array controller functions, including parity 
generation, RAID algorithms, striping
     algorithms, and cache management, freeing up the host system CPU. Each SCSI 
channel gives you up to 20
     MB/sec throughput; a three channel version gives you 60 MB/sec bandwidth. 
Optional EDRAM provides no
     wait-state cache access for faster read and write operations. The cache 
memory is expandable up to 64 MB
     DRAM (8 MB with EDRAM). 
..............................

	Hereafter is an extract from a reply to note 4633
    
    Could someone please help (my customer) with specific question concerning
    the KZPSC-BA RAID controller ?
    
    1. What is the board's aggregate bandwidth ? Is it 20 MBytes per second,
    or is the bandwith of each SCSI channel 20 MB/s making the aggregate
    bandwidth 3 x 20 or 60 MB/s ?
    ...
    Tony
    
>    1. What is the board's aggregate bandwidth ? Is it 20 MBytes per second,
>   or is the bandwith of each SCSI channel 20 MB/s making the aggregate
>    bandwidth 3 x 20 or 60 MB/s ?
 
It's somewhere around 18Mb/s
  ...
  		Hein.

        re .1
    Can I just be sure, is the bandwidth of each SCSI CHANNEL 18 Mb/s, or is the
    overall bandwidth of the KZPSC MODULE 18 Mb/s ? (as our customer
    believes to be the case).
    
    Tony
    
        ...18MB/sec ?
          ^
        
    	 
>    Can I just be sure, is the bandwidth of each SCSI CHANNEL 18 Mb/s, or is
> the overall bandwidth of the KZPSC MODULE 18 Mb/s ? (as our customer
>   believes to be the case).

Your customer is correct. The overall, or aggregrate, or PCI bus, or however
you want to call the 'controller' bandwith is less than 20 Mega Bytes per Sec.
        This is to some extend indepenend of the SCSI channel speeds as reads 
    might be satisfied from cache, and burst of writes can be cached.

hth,
	Hein.

> This is to some extend indepenend of the SCSI channel speeds as reads might be
> satisfied from cache, and burst of writes can be cached.

I've seen a performance paper on the SWXCR (KZESC) that said that the limiting
factor was the backplane on the controller module, and that the peak
figure referred to a pure cache read scenario.

Having to actually reference the disk gave soemwhat lower numbers...

Since the KZPSC is basically a PCI version of the KZESC, I wouldn't be
surprised if the limit was again the controller backplane - not the
PCI<->controller or controller<->SCSI.

-cw
>> This is to some extend indepenend of the SCSI channel speeds as reads might be
>> satisfied from cache, and burst of writes can be cached.
>
>I've seen a performance paper on the SWXCR (KZESC) that said that the limiting
>factor was the backplane on the controller module, and that the peak
>figure referred to a pure cache read scenario.
    
    Yes, that's how I understand it too. I added that paragraph trying
    to pre-empt 'the next question' which I expected to be 'but what
    if I switch/have_set the channels to 10MB/sec, shouldn't I get 30MB/s?
    The answer being: No, You'd still get < 20MB/s.
    With the channels set to lower speeds, and with real disk accesses
    involved like a 'real application' the MB/s will come further down.
    Check for a paper in the IO_PERFORMANCE conference on SSDEVO::.
    
    Hein.      
    
    
    	

    60 MB/sec is the theoretical bandwidth of 3 x 20MB FW SCSI busses ... I
    suspect 18 MB/sec is the measured bandwidth through the controller
    using a specific block size.

    
    	Straight from the horses mouth - Mylex web site.
    This tracks very close to all other entries in various Digital
    notesfiles that say 18 MBytes/sec is max no matter whether PCI  or
    how many channels.
   __________________________________________________________________________
    	As explained before, performance on the PCI controller is very
    dependent on the distribution and quantity of drives per channel. For
    example, only one drive per channel on DAC960-P or 3 drives on the 1st
    channel only give dramatically reduced performance (around 15 MB\s). To 
    achieve 28 MB\s, the DAC960P-3 controller must use EDRAMs and be loaded
    with atleast 2 drives per channel and have atleast 6 drives. Using
    standard DRAMs, will give  only around 16 MB\s. 
   ___________________________________________________________________________
    
    						Bill

    I am not sure if this will help, but we in the PCBU Get querried
    about Mylex data rates also.  We have asked our friends at mylex, and
    the general reply is 'it depends'....what it depends upon are many
    things that need to be understood so the numbers make sense.  I can
    share a mail I got that may help...it only covers RAID0, but heres
    the info;
    

Here is the configuration information that you had requested for your   
RAID0 benchmarking.  Let me know if you have any questions.

RAID 0 configuration recommendations for benchmarking.

Here are the items that impact performance:
The number of drives.
The number of SCSI channels available
Stripe Size
Cacheline Size
Read Ahead
Memory Size and type (EDRAM or DRAM)

WriteBack should be enabled for  benchmarking tests requiring higher   
throughput.  Write Through for minimum latency.

Should use 8 drives split over three SCSI channels to remove the   
bottleneck at the drive level.


If tests are Sequential Reads:
8 MB EDRAM
Read Ahead Enabled
64k stripe
64k cacheline

Maximum IO throughput that can be achieved is 29MB/s


If tests are Random Reads (100% reads):
8MB EDRAM
Set the Cacheline size according to the maximum transfer size.
Most preferable setting is 64KB cacheline, 64KB Stripe  and Read Ahead   
disabled.


Sequential Writes:
Same as Sequential Reads.
Maximum IO throughput that can be achieved is 22MB/s


Random Writes:
Has a lot of drive dependency.  Should use 8 drives, but with very good   
drive seek times.  We have used Seagate Barracudas in our testing.
Set the Cacheline size according to the maximum transfer size.
For large IO transfers performance is dependent on memory size.  Use DRAM   
with greater size to increase performance.
    
    
    I truly hope this helps.
    
    chet white

    
    	Thanks Chet,
    
    	A follow up question would be since we offer the EDRAM only on the
    KZPAC not the KZPSC, can we change the DRAM to EDRAM on the KZPSC?
                                                    
    					Bill

    
    Bill,
    
    I'm 'really' not up on the storageworks terminology....ie?: kzwhatever
    versus the DAC9xx and fr-numbers........
    
    Also don't know what the storageworks officially support for this,
    but since it's a mylex anyhow, I'd check with their webpage, and look
    to see if the module spports it....www.mylex.com  I'll check sometime
    later today.
    
    chet

    Re: 6376.3
    
    The information you provided for RAID 0 benchmarking was very useful. 
    However, my customer needs RAID 1 benchmarking configuration info.  Do
    you have any specific info for optimizing RAID 1 benchmarks?
    
    
    Thanks,
    Olin Holly

    I will make the request......
    
    stay tuned.
    
    chet