From pctr@ing.iac.esMon Oct 21 10:56:10 1996
Date: Wed, 2 Oct 1996 11:52:55 +0000 (WET)
From: "Paul Rees, ING" 
To: Rene Rutten 
Cc: Gary Mitchell , Pete Fishwick 
Subject: Re: more power !

Hi Chaps,

In reply to Rene's email about the performance of lpss1 and lpss3, I have had a look at the systems and can make a few comments:

lpss1 is currently running as an NFS server, network server and mail server, as well as being the first choice for user logins/compute serving. To get better cluster performance it should be either serving NFS, mail and the network and not allowing any user logins, or it should be serving the computing. This is probably the main reason for its percieved user performance
If you want more performance from lpss1 in its existing configuration, its memory could be increased from 92 Mbyte to something more realistic in terms of its role (i.e. greater than 128 Mbyte, if possible).
lpss1 is also the machine on which backups are performed daily. These are done to DAT tape. The two tape drives are connected to the same SCSI controller as the disk drives. This will lower the disk access (and thereby NFS) performance of the machine.
The SCSI bus on all Sparc 10 systems is capable of SCSI 2 performance, with twice the data bandwidth of SCSI 1. However, to perform at SCSI 2 bandwidth the SCSI daisy-chain has to be actively terminated and have only SCSI 2 devices on it. lpss1 has a passive terminator on its SCSI daisy-chain, which means that even if all its peripheral devices were SCSI 2, it would still only turn in SCSI 1 performance. This, again, would lower the disk access and NFS performance.
lpss3 is a Sparc 2 processor with 64 Mbyte of memory. As such, its raw computing performance is rated as around half that of lpss1. It's primary function has evolved into DAS Real-time Display. Running this and data-reduction software on big images is going to turn in pretty poor performance. The solution to all this is not simply to buy more machines, although I do believe we lack a real compute server on the mountain top. Much of the percieved problem with performace may just be lack of memory, which is a well known performance bottle-neck for heavily used Unix machines. With current memory prices being a quarter what they were a year ago, this upgrade path is comparatively cheap and would turn in an immediate improvement. I also agree with Rene' that we need to bid for at least one more machine - my justification would be that if we are to move to ever larger CCDs we need our computing requirements to keep up, otherwise "quick-look" will become a euphamism.
Next year it is intended to replace lpss3 with Daniel's Sparc 20, originally bought for use at sea level for DAS/CCD testing. I do not know how much memory this machine currently has, but it should have at least 128 Mbyte before becoming operational. This will solve the lpss3 performance problem and host the WHT DAS upgrades.
lpss1 is a little more subtle. My money would be on upgrading this machine to 128 Mbyte and adding an additional SCSI 2 Sbus card to support tape drives. Then ensure that no SCSI 1 devices were put on this, or any new, machine.
For the move to Solaris 2, the Sparc 5 allocated for the library would be ideal, in terms of its ready availability and performance, to be the initial Solaris 2 NFS, mail and name server for the mountain-top. We also need to run NIS-plus on the mountain-top rather than NIS, this is a significant change which requires a Sparc 5 minimum to run effectively.
I know that Gary has some of the above in hand and may have replied to Rene' already. I believe we need to discuss detail for the move to Solaris 2 on the mountain-top very soon, because of its operational implications as well as the outstanding need to move to Solaris 2. I would suggest that once we have an outline plan for the move to Solaris 2, a bid is prepared to bring the computing power on the mountain-top up to strength (probably with the purchase of two additional Sparc CPUs).
Cheers,
Paul
PS - I haven't mentioned our crop of IPX and Sparc 2 machines (and one ELC). I believe these machines have no future now except as glorified X-terminals. Sun are unlikely to support these machines for much longer. We'd possibly be better off trading them in for fewer, newer and more powerful machines.
```
  ______________________________________________________________________
  Paul Rees
  Isaac Newton Group                           Phone: +34-(9)22-40 55 00
  Apartado 321                                        +34-(9)22-40 55 68
  Santa Cruz de La Palma                       Fax:   +34-(9)22-40 56 46
  Tenerife, Spain                              Email: pctr@ing.iac.es
  ______________________________________________________________________
```