* Posts by Cesar Maciel

10 publicly visible posts • joined 30 Apr 2009

Intel reveals pay-to-play Xeon features with software-defined silicon

Cesar Maciel

Re: Nothing new here

Mainframes ARE servers, much like in the way of any other server - they use the same Von Neumann architecture. Where they differ is in scalability and resiliency (and multiple embedded accelerators, and the different operating systems, etc), but to call them "not a server" is simply incorrect. Even IBM calls them servers (or, to be more modern, both mainframes and Power and called "Systems").

The ability to ship a full machine with resources deactivated and activate using software keys is part of the mainframe for decades. But every other commercial system has been offering a similar capability for ages. IBM I (former AS/400) had this in the 90s. IBM AIX-based systems (pSeries, System p, Power Systems, depending on the year) introduced this capability in early 2000s. HP Superdome had it. I believe even the Sun Fire 15K had it. It is nothig new, has been old news for ages, except that it is news now because Intel wants to do it. It is news for Intel, not for the IT world.

And yes, it works fine, and it is a way to make a server more affordable. And it does not make it more expensive (maybe slightly if you go for full activation, but without the feature, the manufacturers would simply charge the full price on it).

Moreover, these days the HW cost is not the important factor. If Intel charges, say, $2000 to activate a core, Oracle charges $47,000 to license it to run Oracle Enterprise. So, if no all cores are needed, turning them off brings a significant costs savings. This is the major motivation for the capability. The second one is the instant upgrade that it allows, with zero or near zero downtime.

I worked for IBM on technical pre-sales for Power Systems (still work for IBM in another area). I designed many solutions using on demand capabilities on Power Systems, and several large companies use it regularly, along with dynamic relocation of resources, to manage their production systems.

IBM walks away from £870m T-Systems mainframe deal amid competition concerns

Cesar Maciel

Re: @AC No bullet dodged...

>W.R.T the mainframe... the interesting thing would be to port Hadoop/MapR and Kubernetes to the >mainframe. Now you have a very secure, higher performing base system which can be virtualized >as cloud infrastructure. If IBM were 'smurt' they would be doing this already.

It's been done years ago, dude.. By IBM and also by non-IBM - Veristorm ported Hadoop to the mainframe in the 2014 timeframe.

IBM's own ICP - IBM Cloud Private, it's own Kubernetes environment - runs on Linux on the mainframe.

It is not true that most enterprises are offloading the mainframe - yes, there are applications migrating, but it is not all of them, and some times they are being migrated to leave more space to other applications that are growing. The following link puts an interesting perspective on where the mainframe stands today.

https://blog.syncsort.com/2018/06/mainframe/9-mainframe-statistics/

IBM so very, very sorry after jobs page casually asks hopefuls: Are you white, black... or yellow?

Cesar Maciel

Re: Aren't we becoming too sensitive to everything???

Yeah, I did a quick Google check just before posting, but I did not go further (it wasn't the main point anyway). But thanks for the clarification and more in depth information :-)

Cesar Maciel

Aren't we becoming too sensitive to everything???

I am Brazilian. Mulato and Amarelo (mulatto and yellow) are not racial slurs in Brazil, and we even expect to see these options while picking ethnicity. Also, calling someone Black *is* racist in Brazil! You have to call them negro (which is racist in the US) - or the politically correct and scientifically stupid "afro-descendent" (stupid because humans first appeared in Africa, so technically everybody is afro-descendent).

The tag "Brazil" in the code hints that the form was written in Brazil, and although it would be used internationally (and I completely agree that the code should follow the rules for the deployment location), the issue can very well be attributed to the fact that the brazilian portuguese speaking developer added it as the normal rule for the local convention.

It may be bad code (since it was not meeting the rules of the target audience), but it is really an excess of freaking out to call it racism.

Low AI rollout caused by dumb, fashion-victim management – Gartner

Cesar Maciel

This is a very biased evaluation. If the performance of the AI system is the same as a human, then there is not a motivation to use it (one can argue about cost reduction in the long run, etc, but most companies will not do that unless there are immediate productivity gains also).

However, there are many tasks that are being executed today by AI systems (just using the term so that I don't have to write "machine learning, deep learning" all the time) that are more efficient than humans - computer vision, for specialized activities, is one example (such as image tagging and classification, content tagging, etc). And as mentioned before, cancer diagnostic and treatment has been a sweet spot for AI systems. Using computer vision to identify skin cancer proved to be more accurate than trained oncologists. (see the two posted links below).

In my opinion, dealing with an AI-based "personal interaction" suffers the same problem of fear of the unknown as nuclear energy - most people are against it because they do not comprehend it, and don't care to learn about it.

http://www.nydailynews.com/news/world/ibm-watson-proper-diagnosis-doctors-stumped-article-1.2741857

http://www.kxxv.com/story/38309412/ai-detects-skin-cancer-better-than-dermatologists-in-international-study

Nvidia's Tesla P100 has 15 billion transistors, 21TFLOPS

Cesar Maciel

This adapter is based on NVLink, not PCI-e.

http://www.nvidia.com/object/nvlink.html

Better hot-patches, more virty in Big Blue's next AIX

Cesar Maciel

Re: odd EOLing

Just a clarification - AIX 7.1 Enterprise Edition (the product that brings AIX 7.1 OS, PowerVC, PowerVP, and others) is the one being sunset, not the AIX 7.1 OS. And the reason for the sunset is because some of the products bundled inside AIX 7.1 EE will change, and therefore a new product number is required. AIX 7.1 OS will still be available for a long time, along with AIX 7.2.

Storage upstart: Our flashy gear is WAY faster than slow old DRAM

Cesar Maciel

Seems an idea similar to the IBM POWER8 CAPI Flash solution

The IBM Data Engine for NoSQL, based on POWER8, a PCIe FPGA accelerator from Nallatech that is connected through a coherent interface called CAPI, and the Redis NoSQL database, delivers the same combination of fast DRAM and cheap Flash seen as DRAM. The FPGA maps the Flash as a memory block for the redis database to use, using key value store operations. The CAPI interface allows the accelerator to participate in the coherency domain of the processors - in other words, it can directly access main memory, without requiring system-to-device memory copy that is common in accelerators that are seen as I/O devices.

Since Flash is cheaper than DRAM, you can have lots of memory on small servers at an affordable pricing. The S822L supports two IBM FlashSystem 900 connected as "real memory", for a total of 114 TB of "RAM" - on a 2U system. The Redis SW can mange DRAM and Flash for optimizing response time requirements, and the solution offers an attractive balance between performance, cost, and simplicity.

There is a paper describing CAPI and the NoSQL solution with POWER8 at http://www-03.ibm.com/systems/power/solutions/bigdata-analytics/data-engine-nosql/

Cesar

Disclaimer: I work for IBM. This comment is my own opinion and may not represent IBM's position or statements.

Oracle plops true live migration onto SPARC hypervisor

Cesar Maciel

Max number of partitions on IBM Power Systems

PowerVM was enhanced in April 2011 to support up to 320 partitions on the Power 750, 640 partitions on the Power 770 and 780 and up to 1000 partitions on the Power 795.

IBM slips Power6+ into racks, blades

Cesar Maciel

Minor Corrections

Very nice article, but there are a few incorrect details, as follows:

1 - The Power 520 and 550 have had 2.5" SFF SAS disks for a while, and SSD disks have just been announced, as shown in http://www-03.ibm.com/systems/power/hardware/520/specs.html

2 - The Power 595 has the same 32MB L3 cache per POWER6 chip as the smaller systems, as shown in http://www-03.ibm.com/systems/power/hardware/595/specs.html . What's different between the smaller systems and the 595 is that the 32MB L3 cache on the lower systems is a single 32MB DRAM chip, while on the 595 it is split into two 16MB DRAM chips. That doubles the number of buses used for L3 cache I/O, therefore doubling bandwidth, as described in http://www.research.ibm.com/journal/rd/516/le.html

Regards

Cesar