Free-Writing Bloggerooni

Well, to be honest I think I’ve grown out of the blog.  I was using it as kind of a aide memoir, but I’m now just writing everything down in my book and keeping it in my (falible) memory.  Perhaps I should re-start doing it.

So as I was saying, I went on holiday, and since then never got back into the pattern of updating the site - so here we are 6 weeks later, a holiday down without snapping my legs, a group presentation, handing in my first paper - a speed PhD (!), new house, newly flooded house,  building site of a house - later….

Here’s some musings anyway.  See how far I’ve moved - I’m now back on a more medical theme - thinking about the audience for my proposed PhD work, who are more medical than CS people…

I now believe it feasible to use a low-overhead polling/alerting protocol, similar the management/agent system used by SNMP in packet switched networks to monitor and manage a multiprocessor machine (almost regardless of size #CPU/chips/cors.  Our group to date have not really considered the management of our SpiNNaker machine, and the output directed to the users as the run of the neural network on the machine is underway.  MMK has a Java system written that does health checking of the machine using a simple protocol that he has designed.  He will be going through this with me w/e 13th March.
We looked at the MIB tree structure to store information and navigate it as a concept, and speculated we could use this for monitoring application data as well as system/network data. The SDRAM could be used to store results/intermediate data as there is just not the bandwidth to get everything in/out of the system on the max. 100Mb/s Ethernet links.  (which also has to carry input data into the ANN).
My hypothesis is that this MIB/mgmt system could be used to get all traffic in and out of the machine for all purposes…   loading code, taking stats,  inputing stimulus.
I also outline to my supervisor a hierarchical approach to collection of data/diagnostics for management purposes - taking the 35,000ft (over)view, and then drilling deeper down into the machine to get more specific information, down from System level, to a Chip cluster (eg. 16*16 chip), super node (eg. 4×4 chip), chip level, then down to the processor and it’s registers/addressable memory…  (down to the neuron!?)

We have a relevant metaphor:  due to the limited exterior b/w getting information is like shining a flashlight around a very dark big black box - we can only see little bits at a time.   I am looking to present information to the users in a number of hierarchical ‘views’ for example:
a)  Diagnostics  -  health of the hardware
b)  CPU Resources - How hard it’s being worked/stressed
c)  Spiking activity - also tied into network activity/capacity  (can we run probes within the ANN to examine performance?)
d)  Locality - looking at the # local spikes vs remotely routed events etc.
It would be amazing to see a real-time glowing/modulating view of the machine in operation, similar to the way that fMRI can be used to peer into the human machine and look at activity centres - plus it would keep the medical people who are ‘customers’ - who have commissioned the simulations interested too - Can the patterns being viewed be related to the organic?… and they can change their viewpoint/zoom on demand to focus on anything they deem interesting…
Can we build an adaptation layer - to take information from our own light-weight protocol out of the system, and map this onto a higher level model, eg, TCP/IP <–> raw Ethernet frames,   Physical Chip/Memory Locations <–> MIB tree mapping.  ie. take the complexity out of the machine as far as possible - let it get on with the neural simulation, and leave it to external machines to do the interpretation?

That is all - whether I come back and type more - who knows!   bye!

Uncategorized

Finally got the COMP7000 presentation off my back today, a catalogue of problems on the way though.

Both myself and Michael who was also giving a presentation today turned up at the usual venue, but no-one turned up.  I checked my phone (on silent) and found that the venue had changed, so we went across there to the other side of the building.  Plugged in the laptop and it didn’t detect the monitor, so had to restart - and it’s a fast rebooter!

Michael’s session lasted ages and left me with 25mins on the clock!    Which I duly took as I talked in and around the points there.   So the feedback was about the over-length, and that I could cut out most of the SpiNNaker stuff - or at least the bits that aren’t specifically related to my area of research,  the fact I could have added in referenced to GRID computing too - which was a good pointer for me, but overall the feedback was positive about the presentation & style.  So a good result.

I also thought about looking into SNMP and RMON as a possibility for the multi-chip machines…

Uncategorized

I am using Bye t host for my hosting,  it’s free.

It comes with a few strings, such as if an error is generated then instead of a 404 it gives an advertising page.

Today there was a new one.  If your webpage contains the word ccrackking  (I’ve misspelled this for obvious reasons), then it busts your page.    And when that happens it goes to an advertising page…

So - I’ll be looking for a new hosting provider too,  as apparently I am no longer allowed to use certain words on my website (and this was introduced with no notice either!)  Pathetic and pretty poor form.   Took ages to track this down - their websupport just pointed at the ToC.

Uncategorized

As of 00:30 today - all quintuple items fixed.

With exception of not being able to get VMware to spot it’s network adaptors… rubbish!

Uncategorized

Rubbish day yesterday, just a set of failures one after the other.  Here in no particular order:

1. Being unable to transfer the Humax video into an editable format on my PC.
2. Losing all the channels on my Windows media setup - I thought I’d try to capture the above using UHF so searched for new channels - big mistake - 0 found!
3. Live Fedora10 Linux onto USB stick didn’t work - something I’d seen in PCW that might allow me to boot the silent PC in Edinburgh more quickly!
4. VMware server timing out leaving me unable to use my VMs. Tried downloading but appeared to be corrupt after 1hr, will have to download here in the office
5. Not completing my presentation, due to all of the above

There were some +ve’s though:

1. Getting the Lewis Hamilton programme off the Humax box, installing VLC and being able to play it on the PC (via USB)
2. Finding the appropriate diagrams for my presentation
3. A good game of badminton with the new raquet
4. The blueberry vodka from the Christmas stocking was rather good

Uncategorized

Well that’s me in the office for the first time in 2009. And I feel energised & ready to go.
The few days off were good - although I still feel I never get enough time to myself to just fiddle with stuff and get it working
Anyway, I’ve got 2 tasks to be completed before heading off on the ski holiday at the end of the month.

1. Presentation for COMP7000
2. Testing of Diagnostic registers and counters in SpiNNaker simulation

It must also be down to the fact that I’ve got real deadlines to approach now - the presentation needs to be done by the 16th when I’m giving it (tx to maw for presenting gadget/laser for b’day - that works a treat!), and the rest needs doing before month end - so I’ve got some targets!

Today has been productive, up and out of the house on-time and head down for a good few hours, and I now have my draft presentation on paper and ready to write up. It’s about 15 pages - of which 13 are real slides and not just intro and outro, I think that should be enough for a 15min talk. It’s been difficult to work out what to leave out as there’s just so much draft material, but I think it’s about right!

Oh oh. I’ve just seen Wordpress 2.7 is out, and I think I should upgrade!

Uncategorized

Yet again, I’ve wasted my morning on bloody linux issues.
I ran a security update and restarted the machine. Booted up - NFS fails to start, so I can’t log into the network. Have to log in as root, and sniff about the place to try and find the cause of the error.

rpcinfo firing up an error saying it can’t connect. I eventually track it down : http://forums.fedoraforum.org/showthread.php?t=207495 and more and more searching about finds a workaround which is to set SELinux to permissive. I’ll turn it back when the new version of SElinux is released which supports the new bind version.
(or something).. because unless your a Linux nut, you can’t understand much of the discussion - all I wanted was a way to fix my damn machine - but no, that’d be too easy. Linux is not for desktop users - simple as that!

Good news on Friday though before I headed off for the train in the afternoon - I can now properly work remotely - installed Exceed on my machine and once I’ve VPN’ed through, then I can run the X applications after logging onto one of our machines.
Mobile broadband however makes SoC designed a nightmare to run! Once we eventually get the new house and going to be looking straight away into the wired broadband option!!

Uncategorized

Yesterday was a write off!
Did some reading in the morning, and then headed into town for the APT xmas do. All very civilised, and lots of people there I’d not met before - some Alumni etc.. Was good to chat with a few of them, esp those moved on to Silistix. The afternoon just flew past, and after the meal we headed to a pub in Didsbury - no SpiNNaker representaiton other than me which was bad !
Drank until about 9:30pm and then to an Indian restaurant for meal - finally heading off home just before 11. Can’t believe the day went so quickly! (head good this morning too - must be getting in practice!)

Uncategorized

Cambridge was fun - a chance to mooch about, preach my own type of ‘how I got my PhD place and funding ideas’ and speak with those considering taking up research in it’s many ways. Strangely I’ve kind of motivated myself with all the good stuff about research, and I’m determined to get out of my little hole that I’m in “finding my research has already been done!”
Plus some of the seminars were good, particularly the ones I was kinda forced to go and see as I was looking after the room. A good 3 days - pretty packed, but grand.
I saw a couple of duffers too - a particular one about changing the internet to route on tags and information rather than on packet destination felt very wrong. I kind of see what they’re getting at, but I don’t think this will happen as it’s too hard to retrofit - hell we haven’t even started IPv6 properly. Which is going to end in a nightmare by the way in the next 5 years - we’ll get some crisis stuff in the papers, and loss of functionality as provider-level NAT has to be deployed! (About the tag routing - we kind of do this anyway! - isn’t a Google search just routing performed on keywords a similar ideas?)
Today is about tidying up everything in the office and moving on. Mukaram is going to give me that heads up on running the model again, so I can start looking at testing out the diagnostic registers/counters, and I’m finally getting to the bottom of various inboxes that proved difficult to manage from a mobile with the signal fading in and out in the room at Cambridge.
We’re so close to xmas now, there’s a couple or three half days for me, Friday up to Edinburgh, Monday for team xmas lunch, and probably everyone will go home early on xmas eve - meaning there’s hardly a day that will be a full day. - Woo-hoo!

Uncategorized

I’m just not getting anywhere - we finally got the model (and debugger - now that was the latest thing to crash and burn after getting everything else to work).
So I now have my model - but today I wanted to spend some time with Mukaram on the model walking through the Diag1 test, so he’s come in at 2pm, and I’m off to the radio about 4pm then onto the Friday drink club at the Ducie. We’ll also be looking at exporting the project down to the Verilog model too - which will be interesting indeed! Just needing that kickstart!

I’ve been thinking too about the buffering and routing while going through the various papers I dug out on a binge and printed. The buffering is an issue - as effectively without buffering we can’t offer much in terms of QoS other than a drop or a tx. But we can be clever with the drop algorithm, and link that in with the measurement and reporting type activities.
The routing and switching nomenclature is a bit muddled with the LAN switching stuff. There’s cut-through switching, although that requires a full-buffer from the bits of documentation that I’ve read. Back pressure should let us know whether to tx or to drop this. A great derivative of this and looking promising is wormhole - looks much more exciting as it’s not needing any more buffering than for a flit. (and as the signal paths are typically this narrow).
The issues with buffering I’ve yapped on about already - it’s expensive, in terms of space on the die, and power utilisation and logic provision around it.
For SpiNNaker the delays on packets are truly not important, wrt the CPU timings, but as we’re swopping in and out neurons - effectively doing a TDM on the CPU resource, then it is important to get the packet out!
I’m hoping that when I get back on Thursday we can do some proper emulations and get things moving again.

The more I read the more I think that I can’t find something that hasn’t been done before. Apparently this is very common, and that my niche will come and find me the deeper I get into the topic! I’m beginning to believe that the niche found by adding the words “on the SpiNNaker architecture” might not be satisfying for me, and that I might end up with putting together like a standardised method of instrumenting network on chips by extrapolating from work on SpiNNaker and simulators like ns.

Uncategorized