Content in Intel Communities

TM2C: a Software Transactional Memory for Many-Cores

webadmin@intel.com — Sun, 13 May 2012 16:59:42 GMT

Dear All,

I recently presented at EuroSys'12 our work at EPFL on a Distributed Transactional Memory system tailored for non-coherent many-cores. The title of the paper is "TM2C: a Software Transactional Memory for Many-Cores". Naturally, the SCC was the target platrform for this work.

The paper discusses in details the design of TM2C and our experiences during this work. You can find the paper here: http://dl.acm.org/citation.cfm?id=2168872&dl=ACM&coll=DL.

If someone needs more details about this work, feel free to contact me.

Best,

Vasileios Trigonakis.

RCCE Communication gets stuck

webadmin@intel.com — Mon, 30 May 2011 16:38:15 GMT

For a long time, me and the colleagues working on the marc026 SCC have the following problems:

The barriers (RCCE_barrier) are not always passed.
The communication may get stuck (both with RCCE and iRCCE libraries).

These problems appear when trying to run a program on more than 8 cores. There were times that 12 or even 24 cores ran properly.

I am using the attached application to test how things work (a ring of nodes and a token sent over the ring). I have run tests up to 1 billion token-hops upto 8 cores. More than 8 (or 12) cores it stops. For example, today on 4 cores:

rck00: [0] Started. The token will be resubmitted 100000000 times!
rck00: [0] Token here: 1003
rck00: [0] Token here: 2007
rck00: [0] Token here: 4015
rck00: [0] Token here: 8031
rck00: [0] Token here: 16063
rck00: [0] Token here: 32127
rck00: [0] Token here: 64255
rck00: [0] Token here: 128511
rck00: [0] Token here: 257023
rck00: [0] Token here: 514047
rck00: [0] Token here: 1028095
rck00: [0] Token here: 2056191
rck00: [0] Token here: 4112383
rck00: [0] Token here: 8224767
rck00: [0] Token here: 16449535
rck00: [0] Token here: 32899071
rck00: [0] Token here: 65798143
rck01: [1] ~~Completed. Token here: 100000000

while on 12:

rck00: [0] Started. The token will be resubmitted 100000000 times!
rck00: [0] Token here: 1007
rck00: [0] Token here: 2015
rck00: [0] Token here: 4031
rck00: [0] Token here: 8063
rck00: [0] Token here: 16127
rck00: [0] Token here: 32255
rck00: [0] Token here: 64511

it got stuck at this point.

The analysis I did showed that it should be a flag synchronization problem.

Did/Does anyone else have the same problem or is it just our SCC?

Thanks,

Vasileios

Error code 1

webadmin@intel.com — Mon, 09 May 2011 13:17:54 GMT

Hello,

Does someone know what does the error code 1 mean?

From pssh I found:

1: at least one remote command returned a non-0 value (but all commands ran)

But since the pssh call runs my program only I cannot make an interpretation based on this.

Thanks.

RCCE Shared Memory Allocation

webadmin@intel.com — Fri, 25 Mar 2011 14:27:09 GMT

Hello,

I stepped on the following issue: RCCE_shmalloc allocates one new RCCE_BLOCK (sizeof(RCCE_BLOCK)=12 bytes) in local memory for every shared memory allocation done, thus the local memory needed for allocating a (pointer to a) double for example is more than the actual shared memory space allocated.

The program I attached, tries to allocate 100MB in single chars and fails (as expected, since it would need 1+GB for local metadata):

for (i = 0; i < Bytes; i++) {
        p = (char *) RCCE_shmalloc(sizeof (char));
        if (!p) {
            PRINTD("RCCE_shmalloc returned NULL (%d byte)", i);
     }
}

According to the output, the program fails after it had allocated around 17668KB = 17.67 MB. If my calculations are correct, the allocation till there needs 212MB of local memory. It fails with error code 137 (Segmentation Fault probably).

I tried to figure out if it was the malloc call inside RCCE_shmalloc that returned NULL, but it wasn't (placed if(NULL) check after the RCCE_BLOCKs allocation).

So, my conclusion is that the RCCE_shmalloc is trying to use space used by the OS (when calling the malloc for allocating the RCCE_BLOCK). Does it make sense, or am I missing something?

PS. I am using RCCE 1.0.13, with SHMADD (512MB of shared memory)

RCCE (emulator) and pthreads

webadmin@intel.com — Fri, 11 Mar 2011 10:28:02 GMT

Hello.

I am trying to use phtreads with the RCCE emulator, but the RCCE functions does not seem to work properly within the threads. Are they supposed to? Are they supposed to work on SCC with Linux?

The simplest example I tried was spawning 1 thread per UE that runs the following code:

printf("UE %d: run from a thread\n", RCCE_ue());
pthread_exit(NULL);

The output from each spawned thread is:

UE -1: run from a thread

I attached the source-file.

PS. I know that the intented programming model of SCC is SPMD, but I would like to use pthreads to run a sevice for the "normal" execution.