No luck with this one. I still think that it is not convenient that what people call "first problem" is actually Problem B, but I am getting used to it already.

+1. Please, X or Z would be much more convenient.

Isn't it 500 best placed?

EDIT: Ok, I think I know who asked this question in the Round :p

Wow, ~900 coders have participated, that's consoling :).

I'd bet around 33. Time 3:00:00.

33 points, time 3:00:33

18 pts, 2:00:00

UPD: A mistake I made when calculating the summation LOL, should be 26pts (same on time)

Funny boundary. In the optimistic standings, there are 790 people with 30+, and 791st is 23. Let's see if 26 (B-full + CDE-smalls?) even appears in the standings after testing is over. Of course, your boundary may be exact even if 26 doesn't appear.

I'll bet a score of 31 with time 99:99:99 will be enough to advance, then.

41 points, time 3:45:51

I'll go with the still-optimistic "14 points and a fast time" :)

Top 500 get T-shirts.

Collect same groups of numbers in the same nodes. So for each number there is only one node which sums up counts from all nodes specific to this number. Then each node sends lowest number with count 1 to main node. Main node choses smallest of them (or zero if there are none).

The only challenge seemed to me to get fair distribution across nodes. If you just send number to node "number % node_count" they will have test case where all numbers will go to the same node. I tried "number % 997 % node_count". Not sure if it was enough.

EDIT: It was enough. 11th place — first time in my life in any GCJ round. Reason — algorithmically problems were easy, so even I wasn't challenged algorithmically. The only problem was "Distributed Computing patterns", which I specifically trained for.

I just did distributed sorting -- sort the numbers in the segment and then pick the minimum out of heap of 100 elements and replace it with the next element. That is O(35M * log 100 + 350K * log 350K). It runs for 12 seconds on my PC, but I don't have 100 cores, so I have hopes that it might pass within the time limit. I couldn't test it on the server, since I was getting weird "Rule violation" errors when trying to submit it for A.

For every prefix, I computed the difference between how many crates are in that prefix versus how many crates should be in that prefix. The sums don't exceed 10^18, and we can take the mod at every step.

But at any one time there are at most 10^18 crates in the warehouse?

Aren't there at most 10¹⁸ crates? 10⁹ stacks times 10⁹ crates each.

They said around half an hour.

From questions: " 3:06:14 We are finishing up our remaining judging and we intend to post the final results within the next half hour or so."

42

26

100

I didn't understand this:

Maximum total size of messages a single node can send: 1 GB.

It's almost four times more then all numbers. Looks like they just wanted to say "unlimited". Or trick people to send too large messages between the nodes and get TLE because of that.

150!

15

70

Btw, how to solve it?
39.

49 ACs XD, sorry guys, I won :D

Nice bet, Swistakk! :P

8MB per one message.

Same problem – got Rule Violation for submitting max test for E.large for A. Technically, I should be within the memory limits for A – I was only sending a single message of 2MB from every node.

I was being hit by the same thing. All responses to my questions about this issue denied it was happening :(

I have two identical submissions on Testrun. One attempts to send a single 500 kB message and gets Rule violation. The other sends 50 times a 10 kB message and works.

Does it get wrong answer or time limit?

You mean C-large. Why didn't you test it on the small input? I did this for every problem and found some bugs.

There is a lot of getting used to in the format.

Because it's something new. Not many solved more than 10-20 distributed problems in their life.

BCD were very easy, however I find E really tricky.

The main factors for me were that, first, of course the main difficulty is very different from the main rounds -- algorithmically none of the problems were hard, the problem is just distributing chunks in a way that all nodes do more or less the same amount of work.

Second, it's much harder to tell whether a solution works or not (how many messages can we send? A few big messages are faster than many small messages, but by how much, and how fast is one small message compared to one big message?) The benchmarks help a little, but it's no substitute for definite experience, and Testrun giving rule violations did not help matters. I didn't consider rearranging all values for E because I thought I didn't have time to send that many values.

Also significant is that it's much trickier to code in these rounds, and it's much harder to debug when a problem does happen. For instance, memory limits are a big problem here, and in regular code jam it's a nonissue.

And finally, I do think E was actually tricky.

A large? Or B large?

I got the same error but MinGW worked.

"To make Java work on Windows, you will have to modify paths and commands in build.py. We are working on a version which would work for common Windows configurations. One option is to use MinGW instead of cygwin. Make sure to add MinGW's environment path before cygwin's. Then, "python dcj.py test --source sandwich.cpp --nodes 10" works (because we can't use shell scripts in MinGW)."

It seems this is required for C++ too.

I can't run it in cygwin, but you can run this in windows cmd directly (without that .sh script):

python dcj.py test --source src/sum/code.cpp --nodes 10

I may be wrong, but won't this fail?

                  char c1 = a.get(b.get(i));
                  char c2 = a.get(b.get(i+1));

b.get(i) seems to contain the index of a.get(i) in the entire sequence, so this should be something of the form a.get(b.get(i) - start) where start is the starting position of your block.

I think that "cp crates1.h crates.h" and recompiling is easier than changing include in code and recompiling.

I downloaded and extracted the provided tool. There is a directory dcj_linux. There, I wrote a script:

cp $1 TODO.h
./dcj.sh test --source pro.cpp --nodes $2

and I've copied my templates into a file pro.cpp. Now I can copy my folder when I start a new problem. After copying I must by hand change name TODO.h in a script into e.g. oops.h (problem name). And I also by hand change #include "TODO.h" in my code into e.g. #include "oops.h".

Then, I can run my program with: ./script lolsample1 100 — this one runs my program for input in file lolsample1 on 100 nodes.

I use JHelper with tweaked templates
I put headers as tests, and for each test it just creates header and run the code. (I don't usually change the number of nodes)

Note very convenient, but still I don't have to run each test separately

How do you include message.h without raising an error?

I find most of supplied testcases useless (except as additional way to understand problem statement better). It's usually 3-5 items. For 100 nodes. So I usually just end up writing up some random generated large input. Or not exactly random — like in "crates" I was testing on N piles with N, N-1, N-2 and so own height. In E I was testing with 70,000,000 players distributed over 2 nodes each producting n+1 as number. So I knew correct answer of 1 and could see (by using PARunner statistics) how big messages were sent between nodes.

I create a symbolic link pointing to the test I want to run:

ln -fs 01.h crates.h
dcj test --source main.cpp --nodes 5

Linux rules.

I feel your pain. I've spent a few hours during the practice rounds modifying the dcj script to make it work and I wrote some additional scripts to switch between problems.

You know, practice round was there for a reason.

Can you choose the language as C++ so that it highlights correctly?

Few things wrong:

• You're sending from node 0 to node 0
• "Receive(-1)" on line 57 wth is that?

I think you have problem because of this two lines:

Receive(-1);
ll x = GetLL(node);

You receive data from any node, but get data from specific node. Following code should work

ll source=Receive(-1);
ll x = GetLL(source);