Intel Confirms Decline of Server Giants HP, Dell, and IBM

[Anne & Lynn Wheeler]

Intel Confirms Decline of Server Giants HP, Dell, and IBM
http://www.wired.com/wiredenterprise/2012/09/29853/

long winded recent thread in ibm-main mailing list ...
google archive via usenet gateway
https://groups.google.com/forum/?fromgroups=#!topic/bit.listserv.ibm-main/wtgFuOaNCak

pieces also here
http://www.garlic.com/~lynn/2012m.html#2 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee
http://www.garlic.com/~lynn/2012m.html#3 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee
http://www.garlic.com/~lynn/2012m.html#4 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee
http://www.garlic.com/~lynn/2012m.html#5 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee
http://www.garlic.com/~lynn/2012m.html#6 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee
http://www.garlic.com/~lynn/2012m.html#7 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee
http://www.garlic.com/~lynn/2012m.html#11 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee

current mainframe in the field is z196 mainframe, max configuration is
80processors and 14 system assist processors that goes for $28m and is
rated at 50BIPS. The thread had wandered into applicability of
mainframe for large cloud operation. Basic stable cloud component is
e5-2600 blade. IBM has base list price for e5-2600 of $1815. Depending
on ghz rate there are range of dhrystone BIPS rate for e5-2600
.... including 527BIPs. For pricing standpoint the comparison works
out to be $560,000/BIPS compared to $3.44/BIPS (or difference of
163,000 times). However, major cloud operators have been claiming being
able to do blades for their own megadatacenters at 1/3rd brand name
blades (and doing hundreds of thousands per megadatacenters would
account for uptic in the non-brand-name blade manufacturing).

The other issue brought up in the thread was efficiency of mainframe
i/o operations with their independent processors and channel
architecture. The issue is that old bus&tag was reasonable trade-off
for the 60s, but became less&less optimal as technology advanced
(especially sticking with serialized sequential operations and
end-to-end latency). They moved to fiber-optic "ESCON" which had
faster transfer rate but retained the basic channel architecture.

I've complained numerous times when mainframe channel engineers
started participating in fibre-channel standards meetings 20some years
ago ... wanting to leverage the faster transfer rate of FCS ... but
layering the old-style channel architecture protocol on-top (aka
"FICON", negating a lot of work in underlying FCS for batched,
asynchronous, concurrent i/o operations).

This has basic FICON operations (in z10 time-frame) with peak of
31,000 IOPS.
ftp://ftp.software.ibm.com/common/ssi/sa/wh/n/zsw03059usen/ZSW03059USEN.PDF

Recently they get around to doing a little batching ... calling it
zHPF that gets it up to 92,000 IOPS peak (for z196 & zEC12) per
fibre-channel channel (slightly corresponding to the underlying
original fibre-channel design point).
http://www-03.ibm.com/systems/z/hardware/connectivity/ficon_performance.html

above describes introduction of TCW ... being able to download
multiple CCWs in one transfer for remote execution ... instead of
serialized one at a time ... getting some FICON throughput improvement
for zHPF

The mainframe SSCH has still quite a bit of processing overhead, this
has z196 peaking at 2M IOPS using 104 FICON channels ... with z196 14
system assist processors having theoritical max thruput of 2.2m IOPS
at 100% utilization.
ftp://public.dhe.ibm.com/common/ssi/ecm/en/zsw03169usen/ZSW03169USEN.PDF

This has a emulex fibre-channel for e5-2600 capable of 1M IOPS per
channel
http://www.emulex.com/artifacts/0c1f55d0-aec6-4c37-bc42-7765d5d7a70e/elx_wp_all_hba_romley.pdf

Part of the thread, I describe in 1980 (over 30yrs ago), writing
support for downloading mainframe channel programs for batched remote
execution (much more efficient than serialized, single CCW at a time)
http://www.garlic.com/~lynn/2012m.html#11 Blades versus z was Re: Turn Off Another Light - Univ. of Tennessee

--
virtualization experience starting Jan1968, online at home since Mar1970