I am trying to write a paper on mainframes vs pc's and I am to the
point of performance.  I am trying to illustrate that today's pc's
with technolgies like distributed/grid computing, and clustering can
outperform a mainframe and essentially eliminate the I/O difference
between the two platforms.  Can someone lead to to a site with some
comparisons or statistics?  Thanks a lot for your help in advance.  I
have included what I have so far below:

Performance would probably have been the biggest difference between
the mainframe and today’s PC’s; however, with processor speeds
climbing and new technology such as clustering and grid computing, PCs
will put up a good fight.  The only big difference I can determine is
the I/O capability between the two platforms.

From researching mainframes, it seems as though they have a lot of
redundancy and serviceability features built in that allow them to
handle I/O much more efficiently than a lower end PC.  A mainframe can
handle a large amount of simultaneous I/Os and has the ability to page
them quickly.  The following chart will illustrate the I/O capability
of a mainframe class machine, just between different sun machines. 
The Sun Fire 15k can have up to 72 I/O slots with a combined bandwidth
of 21.6 GB/sec (www.sun.com/servers)!
Product Specifications	Sun Fire 3800	Sun Fire 4800	Sun Fire 4810	Sun
Fire 6800	Sun Fire 12K	Sun Fire 15K
CPUs	2-8	2-12	2-12	2-24	4-52	4-106
Processor	UltraSPARC III Cu	UltraSPARC III Cu
Clock speed	750-MHz, 900-MHz or 1.05-GHz	900-MHz or 1.05-GHz
Uniboard CPU/Memory Boards	2	3	3	6	Up to 9	Up to 18
I/O Slots (cPCI/PCI)	12 cPCI	16 PCI or 8 hot-swappable cPCI	32 PCI or
16 hot-swappable cPCI	36 PCI hot-swappable	72 PCI hot-swappable
I/O Bandwidth	2.4 GB per second per PCI/cPCI assembly	10.8 GB/sec	21.6
GB/sec
Those are pretty amazing figures especially when an Intel Xeon
platform PC can provide a maximum of 3.2 GB/sec (www.intel.com).

---

Clarification of Question by jkrech17-ga on 28 Sep 2002 11:46 PDT

I am not really familiar with mainframes so I guess what I am looking
for is proof that a PC Cluster can outperform a mainframe at a lower
cost.

---

Request for Question Clarification by voyager-ga on 28 Sep 2002 16:13 PDT

Hi jkrech17,

do you just want us to provide you with an example where conventional
PCs can outperform a mainframe or do you want us to show that PCs in a
certain configuration are always faster than a mainframe?

The second point would be impossible because it is untrue. Mainframes
today are usually geared towards a very special task. There are some
tasks where different version of networked PCs are equal or even
better suited. They certainly are cheaper. However, other tasks still
run much better on mainframes, even at their steep price.

Do you maybe want us to research which tasks are better left to
mainframes and which are well adaptable to other solutions? (This task
might require raising the price tag though, as it would involve the
need to cover various areas of research)

voyager-ga

---

Clarification of Question by jkrech17-ga on 28 Sep 2002 16:33 PDT

I guess I am just trying to determine the difference between the
overall performance of the mainframe vs a certain setup (clustering)
with the PC's.  My whole goal for this section is to just show that
clustered PC's can perform almost as good as a mainframe without
having to spend a couple million dollars.  To me it seems like
mainframes are still useful for things like the weather; however, most
tasks can now be completed without having to purchase a mainframe. 
For instance, what if you are a small company and can't afford a
mainframe?  Even though you may have some data intensive tasks, can't
a group of PC's complete them, even if it is a little slower than a
mainframe can do?  Maybe I just need a few websites showing a case
study or something of a cluster of PC's vs. a mainframe.  If you need
any more clarification, let me know!  Thanks a lot!

---

Clarification of Question by jkrech17-ga on 28 Sep 2002 16:40 PDT

Also, from what I understand, in most case, companies are buying the
modern mainframes to run databases, not for storage. In the old day,
mainframes hooked up to EMC and Hitachi storage arrays can provide
terabytes of storage. But today, with NAS (Network Attached Storage),
or SAN,
even your little Pentium can have Terabyte of storage. 
 
Therefore, storage isn't a real reason to have mainframes anymore. But
database is a very critical part of n-tier architecture. I am very
inexperienced with mainframes and I am trying to show the difference
between a pc and mainframe.  Is there any place that shows some
examples in terms of numbers?  I already have the cost comparisons but
nothing on performance.

Having worked in the IT department at a large bank and at a Top 10
investment firm, I can tell you this:

The reason mainframes still exist at many large companies is because
PC technology is not currently (and may never be) able to do the kind
of transaction processing required by these firms. Most firms of this
type run a nightly batch cycle to "apply" the day's (thousands or even
millions of) transactions to their huge database. They do this because
this is still the best way to process this kind of transaction volume
in the timeframe available, and to maintain such large databases
(databasii?)

The statement you made about mainframes' superior I/O handling is
generally true. While most firms with such extremely large processing
requirements, now that Y2k (Year 2000 bug remediation) is a thing of
the past, are converting to or adding on a PC-based "front-end" system
that supports on-line banking using PCs, their "back-end" system is
still a mainframe that supplies the sheer "transaction crunching"
power required.

Bear in mind that, since customers tend to be very touchy about things
like money, in the event of a power failure or catastrophic hardware
or software failure, these businesses need to be 1) absolutely sure
that they will not lose any data from previously processed
transactions and 2) able to get up and running again immediately (or
as soon as possible). Hence the need for double- and even
triple-redundant systems, and the need to write each transaction 2 or
3 times (high I/O) to support that redundancy.

Consequently, any cost-benefit and performance analysis of mainframes
vs. PCs is dependent on the volume and the timeframe of the required
processes. You will not be able to write a paper showing that PCs can
ALWAYS provide the same performance for a lesser cost. However, you
should be able to make a case that for many businesses, PCs would be
able to provide the performance required at a substantial savings
(remember that the businesses' potential growth factor must also be
considered).

While the information you get would obviously be slanted to the
company's benefit, companies such as IBM, HP, and Sun would probably
be able to provide you with some of the promotional information they
have already developed to persuade potential clients that their system
is the best solution. You might try contacting the sales departments
of these companies, explain what you are looking for, and see if they
are willing to provide with such information on- or off-line.

Hope this information is of some assistance. Good luck!

Hello jkrech17,

I can provide a counter example (and won't post as an answer unless
you want it) that shows that the application may require significant
changes when moving from a large computer (or mainframe) to a cluster
of smaller ones. This is a real example - taken from a system I am
working on to move from an SGI Challenge XL to a cluster of PC's
connected by Scalable Coherent Interface (operating similar to shared
memory).

The Everest bus on the XL (several years ago) provides several
gigabytes per second transfer rate between CPU cards, memory, and I/O
cards. Accesses to memory are cache coherent - a write by one CPU to
memory will update caches for all other CPU's as necessary. A new
Origin machine was priced at about $500k. Go to
  http://www.sgi.com/
for more information. The SGI machines are not typically called
mainframes, but these machines happen to share some of the
characteristics of a mainframe.

SCI has 1 Gbyte/second connections, but data transfers are usually
restricted to 100 Mbyte/sec on 33 Mhz/32 bit PCI (or up to 200-400
Mbyte/sec on 64 bit PCI). Latency for a memory access on another node
is usually 2.5 microseconds, but caching will improve performance to
bus rates on reads / writes to consecutive addresses. Caches across
the cluster are not coherent, and the order of writes are not
guaranteed unless cache flushes are used to sequence writes. The
cluster with a lot more CPU power and memory runs about $80k. Go to
  http://www.dolphinics.com/
for information about SCI, and references to vendors that will sell
SCI clusters.

The application generally works except for...
 - congestion can occur on a bus to delay SCI transfers to over 256
microseconds (measured). This cause problems on a VME bus where an 80
microsecond time out was programmed into the bus controller.
 - the lack of write ordering and cache coherency adds complexity to
code that must use cluster shared memory. Do not under estimate this
latter effect, search with phrases such as "cache coherency", "weak
memory model", or "write ordering" for articles that describe this.

To give you an idea of the impact of this, transfer of 64 bytes in a
message takes about 9 microseconds to move the data and 40
microseconds to get a lock, update pointers, and release the lock.
There is also a lot more code required to do this properly.

Don't get me wrong, we are happy with the new hardware - especially
the cost of maintenance and performance of each system in the cluster,
but the software costs and latent bugs due to a brittle design have
pretty much eaten up the benefits.
  --Maniac