January 2013 - discuss - mailman3.cels.anl.gov

MPICH migration to git
by Dave Goodell 08 Jan '13

08 Jan '13

MPICH users & developers, The MPICH project has transitioned from using SVN as our version control system (VCS) to using git instead. Details of the switch are described here: http://wiki.mpich.org/mpich/index.php/Git Highlights from the document and the transition: * Read-only git clones are available at http://git.mpich.org/mpich.git and git://git.mpich.org/mpich.git * All critical history from SVN has been imported into the git history. * The SVN server and its associated history are not going away, although it has been made read-only. We continue to work through our documentation and automated processes (cron jobs, etc.) to update them to the new system. Our trac instance will transition to understand the new VCS soon, although it currently displays only historical SVN data. Please bear with us as we work through this transition period. If you have trouble with the new setup, please contact devel(a)mpich.org or discuss(a)mpich.org for assistance. Regards, The MPICH Team

1 0

Re: [mpich-discuss] low performance in an asynchronous, mixed MPI/pthreads app
by Geoffrey Irving 08 Jan '13

08 Jan '13

On Mon, Jan 7, 2013 at 11:39 AM, Darius Buntinas <buntinas(a)mcs.anl.gov> wrote: > Hi Geoff, > > Is it possible for you to try this with fewer worker threads per node (leaving some idle cores on each node)? Yes, this should have been an easy thing to try, but it crashed due to passing a 0 pointer to MPI. That may indicate that I ran out of memory and an unchecked malloc failed. I'll track that down and retry. > What's the scale of the x-axis? (about how much time is one tick?) The scale is shown on the top next to the ticks. It's 1e-2 s = .01 s in both figures. > Do the worker threads receive messages, or is it only the comm thread receive all messages? The comm thread receives all messages. > How many isends are performed by the worker in the "wakeup" section? Just one 8 byte Isend. > Aside from the isends in "wakeup," are there any other MPI calls made by the worker threads? No. > How many Isends are performed by the comm thread in "response_send," and how many in "output_send?" It's usually 1 per response_send region, but can be more if two happen to occur back to back (they're separate blocks but you can't tell in the screenshot without selecting one of them). In output_send several occur at once (usually 10 or 20 I think in the figure). The scatter pattern in the lines towards the right of the following shows one example. > Does "wait" do any MPI other than MPI_Waitsome? No, just MPI_Waitsome. > Does the worker do any MPI calls other than isends in "response_send" and "output_send," and waitsome in "wait?" No for output_send and wait, but yes for response_send: after the data Isend I repost a wildcard Irecv for a small 8 byte message. Thanks for all the questions! Geoffrey

1 0

Re: [mpich-discuss] low performance in an asynchronous, mixed MPI/pthreads app
by Geoffrey Irving 08 Jan '13

08 Jan '13

On Mon, Jan 7, 2013 at 10:53 AM, Dave Goodell <goodell(a)mcs.anl.gov> wrote: > On Jan 6, 2013, at 8:57 PM CST, Geoffrey Irving wrote: > >> We're seeing much lower >> performance than expected, and are trying to understand why. Here is >> a visualization of a 96 core test run using 16 ranks with 6 threads >> per core (4 Hopper nodes): >> >> http://naml.us/random/pentago/history-random17.png >> >> 1. All messages are sent using MPI_Isend, but Isend sometimes takes a >> very long time to return even for tiny (8 byte) messages. For >> example, the light blue "wakeup" messages are sent from a worker >> thread to the master communication thread to indicate that computation >> is complete, breaking the communication thread out of an MPI_Waitsome >> call (using MPI_THREAD_MULTIPLE). Some of the input request Isends >> also take a long time, and these are also 8 bytes. What would cause >> MPI_Isend to take a long time to return, both for small messages >> (wakeups and input requests) and for large messages (input responses >> and output sends)? > > One possibility is because MPICH-nemesis (on which Cray's MPI for XE6 is based) does not have any fine-grained parallelism in the library. So in order for an Isend-ing thread to properly enter the MPI library, it must queue on a mutex until the thread which is currently occupying the progress engine decides to yield. In nemesis by default this is 10 iterations of {polling shared memory then the network module}, although Cray may have decided to tweak this number (I have not seen their source). So overly long netmod polling times could contribute to this delay. > > This design is simplistic, and we aren't especially happy with it, but so far we've found it to be very difficult to retrofit finer-grained parallelism into the existing nemesis codebase. My personal opinion is that a real fix will require a substantial rewrite of a large part of the lower level code. > > If I'm understanding your questions/explanations correctly, then this is likely related to issues 2 & 3 as well, possibly 4. > > […] >> Some things that might contribute to the problem: >> >> 1. Since the 92K input responses are sent in response to 8 byte >> request messages, I can prepost all matching Irecvs. However, output >> messages arrive unexpectedly, so the best I can do is post a few >> wildcard Irecvs (MPI_ANY_SOURCE, MPI_ANY_TAG on the output message >> communicator). Jeff thought wildcard Irecvs might not be good enough >> to hit the fast path. Moreover, output messages are not compressed, >> so most are 262K. I didn't compress them historically because >> compression isn't all that fast (I'm using Google's Snappy plus domain >> specific preconditioning), but that may be a good experiment to run. > > Preposted any_source+any_tag irecvs should be on the fast-enough path. The size is large enough that it might be tipping you over into a rendezvous instead of eager protocol, but only Cray can answer that for sure. I probably wouldn't pursue compression as a first next step. And you're correct: at least in the current setup output compression slows things down, probably in part because it results in more intrarank messages (I had already started implementing this experiment before I got your email). >> 2. I don't know how to visualize instantaneous bandwidth, even though >> I have start and end times for all large messages (start and end of >> Isend, Irecv request completion time). It's possible the problem is >> bad network load balancing. I'm happy to whip up a plot if anyone has >> suggestions as to what to draw. > > I'm not sure what the goal is here. 262 KiB messages don't strike me as large enough to worry much about bandwidth on a modern Cray machine. Latencies seem like a more useful metric still, esp. if you aren't seeing much variation as a function of message size. > >> 3. Waking a communication thread out of an MPI_Waitsome is somewhat of >> a hack. It's the only time the worker threads touch MPI; otherwise I >> could use MPI_THREAD_FUNNELED. However, the only alternative I know >> is for the communication thread to poll on MPI_Testsome and check for >> thread completion in between. Might that be better (it's an easy >> experiment to run)? > > Frankly, it might, because of the aforementioned design. What non-MPI mechanism for inter-thread signaling are you planning on using? All the thread signalling is via pthread spinlocks and the occasional lockless spin when I can get away with it. In the MPI_Testsome case the communication thread would check a flag for inter-thread messages without grabbing a lock, then grab a spinlock and recheck. I'll implement this tomorrow and see how it fares. Thanks, Geoffrey

1 0

Re: [mpich-discuss] low performance in an asynchronous, mixed MPI/pthreads app
by Darius Buntinas 07 Jan '13

07 Jan '13

Hi Geoff, Is it possible for you to try this with fewer worker threads per node (leaving some idle cores on each node)? What's the scale of the x-axis? (about how much time is one tick?) Do the worker threads receive messages, or is it only the comm thread receive all messages? How many isends are performed by the worker in the "wakeup" section? Aside from the isends in "wakeup," are there any other MPI calls made by the worker threads? How many isends are performed by the comm thread in "response_send," and how many in "output_send?" Does "wait" do any MPI other than MPI_Waitsome? Does the worker do any MPI calls other than isends in "response_send" and "output_send," and waitsome in "wait?" Thanks, -d On Jan 6, 2013, at 8:57 PM, Geoffrey Irving wrote: > Hello, > > With help from Jeff Hammond and Jed Brown, I am trying to optimize and > scale a solver for the board game pentago on the Cray XE6 machine > Hopper and NERSC. The code mixes MPI and pthreads, and has an > asynchronous message passing structure. We're seeing much lower > performance than expected, and are trying to understand why. Here is > a visualization of a 96 core test run using 16 ranks with 6 threads > per core (4 Hopper nodes): > > http://naml.us/random/pentago/history-random17.png > > Each rank has 1 dedicated communication thread doing almost all of the > MPI calls, and 5 worker threads doing computation. White (of which > there is too much) indicates an idle thread. The color says what's > happening on each thread, and the lines indicate information flow: > ranks request input data from other ranks, respond to input requests > with data, compute once data arrives, send output to other ranks, and > process output for final storage. The compute is yellow and the > processing is interleaved red/blue. The weirdness is as follows: > > 1. All messages are sent using MPI_Isend, but Isend sometimes takes a > very long time to return even for tiny (8 byte) messages. For > example, the light blue "wakeup" messages are sent from a worker > thread to the master communication thread to indicate that computation > is complete, breaking the communication thread out of an MPI_Waitsome > call (using MPI_THREAD_MULTIPLE). Some of the input request Isends > also take a long time, and these are also 8 bytes. What would cause > MPI_Isend to take a long time to return, both for small messages > (wakeups and input requests) and for large messages (input responses > and output sends)? > > 2. There's a common pattern in the visualization where a bunch of > output processing happens immediately *before* computation. Whenever > this happens the output and compute are unrelated in terms of > dataflow. This is probably the same problem as (1). What seems to be > happening is that an MPI_Isend on the communication thread is taking a > long time to return, prevent the communication thread from noticing > that other messages have arrived (the communication will call > MPI_Waitsome once the Isend returns. If both output messages from > processing and input messages for compute stack up, the code processes > the output first in order to deallocate certain memory as soon > possible. > > 3. Stalled wakeup Isends seem to complete at the same time as input > response Isends on the same rank. Is the MPI_THREAD_MULTIPLE support > problematic? > > 4. The message latency is quite high. For example, one of the input > responses takes over .3 s in the above image. Each such response > message is about 92K. It's possible that line is actually more than > one input response from the same rank (I do not merge them currently), > but in that case I'd expect them to complete at different times. > Unfortunately, my current visualization doesn't do a good job of > visualizing instantaneous bandwidth: I have the data but don't know > quite what to measure. Bandwidth averaged over 10s intervals is shown > here: > > http://naml.us/random/pentago/bandwidth-10s-async-random17-nonblock.svg > > I've also tried asynchronous progress mode (6 threads per rank, 1 > progress thread for MPI, 1 communication thread for me, and 4 worker > threads). This gave a 12% speedup but didn't change the picture > qualitatively (apologies for the mismatched colors and different > horizontal scale): > > http://naml.us/random/pentago/history-async-random17.png > > Some things that might contribute to the problem: > > 1. Since the 92K input responses are sent in response to 8 byte > request messages, I can prepost all matching Irecvs. However, output > messages arrive unexpectedly, so the best I can do is post a few > wildcard Irecvs (MPI_ANY_SOURCE, MPI_ANY_TAG on the output message > communicator). Jeff thought wildcard Irecvs might not be good enough > to hit the fast path. Moreover, output messages are not compressed, > so most are 262K. I didn't compress them historically because > compression isn't all that fast (I'm using Google's Snappy plus domain > specific preconditioning), but that may be a good experiment to run. > > 2. I don't know how to visualize instantaneous bandwidth, even though > I have start and end times for all large messages (start and end of > Isend, Irecv request completion time). It's possible the problem is > bad network load balancing. I'm happy to whip up a plot if anyone has > suggestions as to what to draw. > > 3. Waking a communication thread out of an MPI_Waitsome is somewhat of > a hack. It's the only time the worker threads touch MPI; otherwise I > could use MPI_THREAD_FUNNELED. However, the only alternative I know > is for the communication thread to poll on MPI_Testsome and check for > thread completion in between. Might that be better (it's an easy > experiment to run)? > > Finally, a few details about the computation structure follow. I've > tried to make the above independent, so feel free to stop here. > > Thanks for any thoughts! > Geoffrey > > --------------------------------------- > > Values of pentago positions with n stones depend only values of > positions with n+1 stones, so the computation is embarrassingly > parallel within each such "level". However, it is also memory > constrained: holding level 24 and 25 in memory at the same time (one > input level and one output level) requires about 80 TB. The state > space of each level is divided into "blocks", and the space of > computation required to go from level n to n+1 is divided into > "lines". Each line has a set of input blocks at level n and a set of > output blocks at level n+1. Most blocks have size 64 * 8**4 = 262144 > bytes uncompressed. > > We partition all blocks and lines amongst the ranks, then operate as follows: > 1. When a rank has enough memory to allocate a new line, it posts > Irecvs and Isends 8 byte request message to all ranks which own an > input block. > 2. When a rank receives a request, it replies using an Isend from a > flat buffer (doing O(1) work otherwise). These messages are > compressed, and take up around 92K. > 3. Once all input blocks arrive, we compute the line's output blocks. > Each block at level n+1 is computed from several (4) lines, so if > necessary we send the output blocks to the rank which owns them. > These messages are uncompressed, and are usually 262K. > > Code is here for reference: > > https://github.com/girving/pentago > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss

1 0

Re: [mpich-discuss] low performance in an asynchronous, mixed MPI/pthreads app
by Dave Goodell 07 Jan '13

07 Jan '13

On Jan 6, 2013, at 8:57 PM CST, Geoffrey Irving wrote: > We're seeing much lower > performance than expected, and are trying to understand why. Here is > a visualization of a 96 core test run using 16 ranks with 6 threads > per core (4 Hopper nodes): > > http://naml.us/random/pentago/history-random17.png > > 1. All messages are sent using MPI_Isend, but Isend sometimes takes a > very long time to return even for tiny (8 byte) messages. For > example, the light blue "wakeup" messages are sent from a worker > thread to the master communication thread to indicate that computation > is complete, breaking the communication thread out of an MPI_Waitsome > call (using MPI_THREAD_MULTIPLE). Some of the input request Isends > also take a long time, and these are also 8 bytes. What would cause > MPI_Isend to take a long time to return, both for small messages > (wakeups and input requests) and for large messages (input responses > and output sends)? One possibility is because MPICH-nemesis (on which Cray's MPI for XE6 is based) does not have any fine-grained parallelism in the library. So in order for an Isend-ing thread to properly enter the MPI library, it must queue on a mutex until the thread which is currently occupying the progress engine decides to yield. In nemesis by default this is 10 iterations of {polling shared memory then the network module}, although Cray may have decided to tweak this number (I have not seen their source). So overly long netmod polling times could contribute to this delay. This design is simplistic, and we aren't especially happy with it, but so far we've found it to be very difficult to retrofit finer-grained parallelism into the existing nemesis codebase. My personal opinion is that a real fix will require a substantial rewrite of a large part of the lower level code. If I'm understanding your questions/explanations correctly, then this is likely related to issues 2 & 3 as well, possibly 4. […] > Some things that might contribute to the problem: > > 1. Since the 92K input responses are sent in response to 8 byte > request messages, I can prepost all matching Irecvs. However, output > messages arrive unexpectedly, so the best I can do is post a few > wildcard Irecvs (MPI_ANY_SOURCE, MPI_ANY_TAG on the output message > communicator). Jeff thought wildcard Irecvs might not be good enough > to hit the fast path. Moreover, output messages are not compressed, > so most are 262K. I didn't compress them historically because > compression isn't all that fast (I'm using Google's Snappy plus domain > specific preconditioning), but that may be a good experiment to run. Preposted any_source+any_tag irecvs should be on the fast-enough path. The size is large enough that it might be tipping you over into a rendezvous instead of eager protocol, but only Cray can answer that for sure. I probably wouldn't pursue compression as a first next step. > 2. I don't know how to visualize instantaneous bandwidth, even though > I have start and end times for all large messages (start and end of > Isend, Irecv request completion time). It's possible the problem is > bad network load balancing. I'm happy to whip up a plot if anyone has > suggestions as to what to draw. I'm not sure what the goal is here. 262 KiB messages don't strike me as large enough to worry much about bandwidth on a modern Cray machine. Latencies seem like a more useful metric still, esp. if you aren't seeing much variation as a function of message size. > 3. Waking a communication thread out of an MPI_Waitsome is somewhat of > a hack. It's the only time the worker threads touch MPI; otherwise I > could use MPI_THREAD_FUNNELED. However, the only alternative I know > is for the communication thread to poll on MPI_Testsome and check for > thread completion in between. Might that be better (it's an easy > experiment to run)? Frankly, it might, because of the aforementioned design. What non-MPI mechanism for inter-thread signaling are you planning on using? -Dave

1 0

low performance in an asynchronous, mixed MPI/pthreads app
by Geoffrey Irving 06 Jan '13

06 Jan '13

Hello, With help from Jeff Hammond and Jed Brown, I am trying to optimize and scale a solver for the board game pentago on the Cray XE6 machine Hopper and NERSC. The code mixes MPI and pthreads, and has an asynchronous message passing structure. We're seeing much lower performance than expected, and are trying to understand why. Here is a visualization of a 96 core test run using 16 ranks with 6 threads per core (4 Hopper nodes): http://naml.us/random/pentago/history-random17.png Each rank has 1 dedicated communication thread doing almost all of the MPI calls, and 5 worker threads doing computation. White (of which there is too much) indicates an idle thread. The color says what's happening on each thread, and the lines indicate information flow: ranks request input data from other ranks, respond to input requests with data, compute once data arrives, send output to other ranks, and process output for final storage. The compute is yellow and the processing is interleaved red/blue. The weirdness is as follows: 1. All messages are sent using MPI_Isend, but Isend sometimes takes a very long time to return even for tiny (8 byte) messages. For example, the light blue "wakeup" messages are sent from a worker thread to the master communication thread to indicate that computation is complete, breaking the communication thread out of an MPI_Waitsome call (using MPI_THREAD_MULTIPLE). Some of the input request Isends also take a long time, and these are also 8 bytes. What would cause MPI_Isend to take a long time to return, both for small messages (wakeups and input requests) and for large messages (input responses and output sends)? 2. There's a common pattern in the visualization where a bunch of output processing happens immediately *before* computation. Whenever this happens the output and compute are unrelated in terms of dataflow. This is probably the same problem as (1). What seems to be happening is that an MPI_Isend on the communication thread is taking a long time to return, prevent the communication thread from noticing that other messages have arrived (the communication will call MPI_Waitsome once the Isend returns. If both output messages from processing and input messages for compute stack up, the code processes the output first in order to deallocate certain memory as soon possible. 3. Stalled wakeup Isends seem to complete at the same time as input response Isends on the same rank. Is the MPI_THREAD_MULTIPLE support problematic? 4. The message latency is quite high. For example, one of the input responses takes over .3 s in the above image. Each such response message is about 92K. It's possible that line is actually more than one input response from the same rank (I do not merge them currently), but in that case I'd expect them to complete at different times. Unfortunately, my current visualization doesn't do a good job of visualizing instantaneous bandwidth: I have the data but don't know quite what to measure. Bandwidth averaged over 10s intervals is shown here: http://naml.us/random/pentago/bandwidth-10s-async-random17-nonblock.svg I've also tried asynchronous progress mode (6 threads per rank, 1 progress thread for MPI, 1 communication thread for me, and 4 worker threads). This gave a 12% speedup but didn't change the picture qualitatively (apologies for the mismatched colors and different horizontal scale): http://naml.us/random/pentago/history-async-random17.png Some things that might contribute to the problem: 1. Since the 92K input responses are sent in response to 8 byte request messages, I can prepost all matching Irecvs. However, output messages arrive unexpectedly, so the best I can do is post a few wildcard Irecvs (MPI_ANY_SOURCE, MPI_ANY_TAG on the output message communicator). Jeff thought wildcard Irecvs might not be good enough to hit the fast path. Moreover, output messages are not compressed, so most are 262K. I didn't compress them historically because compression isn't all that fast (I'm using Google's Snappy plus domain specific preconditioning), but that may be a good experiment to run. 2. I don't know how to visualize instantaneous bandwidth, even though I have start and end times for all large messages (start and end of Isend, Irecv request completion time). It's possible the problem is bad network load balancing. I'm happy to whip up a plot if anyone has suggestions as to what to draw. 3. Waking a communication thread out of an MPI_Waitsome is somewhat of a hack. It's the only time the worker threads touch MPI; otherwise I could use MPI_THREAD_FUNNELED. However, the only alternative I know is for the communication thread to poll on MPI_Testsome and check for thread completion in between. Might that be better (it's an easy experiment to run)? Finally, a few details about the computation structure follow. I've tried to make the above independent, so feel free to stop here. Thanks for any thoughts! Geoffrey --------------------------------------- Values of pentago positions with n stones depend only values of positions with n+1 stones, so the computation is embarrassingly parallel within each such "level". However, it is also memory constrained: holding level 24 and 25 in memory at the same time (one input level and one output level) requires about 80 TB. The state space of each level is divided into "blocks", and the space of computation required to go from level n to n+1 is divided into "lines". Each line has a set of input blocks at level n and a set of output blocks at level n+1. Most blocks have size 64 * 8**4 = 262144 bytes uncompressed. We partition all blocks and lines amongst the ranks, then operate as follows: 1. When a rank has enough memory to allocate a new line, it posts Irecvs and Isends 8 byte request message to all ranks which own an input block. 2. When a rank receives a request, it replies using an Isend from a flat buffer (doing O(1) work otherwise). These messages are compressed, and take up around 92K. 3. Once all input blocks arrive, we compute the line's output blocks. Each block at level n+1 is computed from several (4) lines, so if necessary we send the output blocks to the rank which owns them. These messages are uncompressed, and are usually 262K. Code is here for reference: https://github.com/girving/pentago

1 0

Re: [mpich-discuss] mpich 3.0.1 won't build shared libs without fortran?
by Dave Goodell 03 Jan '13

03 Jan '13

On Jan 3, 2013, at 8:55 AM CST, Eric A. Borisch wrote: > I am having difficulty getting mpich 3.0.1 to build shared libs when I'm not using fortran; it still runs checks (against false) for the fortran compilers (even though I am requesting --disable-fc --disable-f77) and associated tools and predictably fails those tests, needlessly disabling the "build shared libs" var. > > Any thoughts? I've put the config.log [1] up on pastebin rather than attaching the half-meg file. You can see things go south in lines 498-579. I can reproduce this locally. We were trying to be clever in [2], but it seems to have backfired into preventing shared linking in this case. It looks like the right way to communicate this sort of thing to libtool is to set these variables to "no" instead. I've committed a fix along these lines in [3]. Thanks for reporting this. -Dave [2] http://trac.mpich.org/projects/mpich/changeset/10622 [3] http://trac.mpich.org/projects/mpich/changeset/10820

1 0

Re: [mpich-discuss] MPICH2 version number
by Jayesh Krishna 03 Jan '13

03 Jan '13

Hi, Did you try checking information about MPICH2 from the Control Panel (Add Remove Programs - Select MPICH2 - Click on "Click here for Support Information")? Regards, Jayesh ----- Original Message ----- From: "Pavan Balaji" <balaji(a)mcs.anl.gov> To: discuss(a)mpich.org Sent: Thursday, January 3, 2013 12:01:57 AM Subject: Re: [mpich-discuss] MPICH2 version number On 01/02/2013 01:22 PM US Central Time, Guanhua Zhang wrote: > Once I installed MPICH2 under Windows. How can I know which version > installed? I tried "mpiexec -version". It did not work. Did you try mpich2version? -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji _______________________________________________ discuss mailing list discuss(a)mpich.org To manage subscription options or unsubscribe: https://lists.mpich.org/mailman/listinfo/discuss

1 0

mpich 3.0.1 won't build shared libs without fortran?
by Eric A. Borisch 03 Jan '13

03 Jan '13

I am having difficulty getting mpich 3.0.1 to build shared libs when I'm not using fortran; it still runs checks (against false) for the fortran compilers (even though I am requesting --disable-fc --disable-f77) and associated tools and predictably fails those tests, needlessly disabling the "build shared libs" var. Any thoughts? I've put the config.log [1] up on pastebin rather than attaching the half-meg file. You can see things go south in lines 498-579. Hopefully I've just missed something obvious... (This was building with clang; similar happens with llvm-gcc.) Thanks, Eric [1] http://pastebin.com/snr3VGdf

1 0

Re: [mpich-discuss] Hard-coded path to mpiexec.hydra.conf (1.4, 1.5)
by Pavan Balaji 03 Jan '13

03 Jan '13

On 01/02/2013 01:25 PM US Central Time, Yauheni Zelenko wrote: > I think this could be solution, but as it looks like also source code > and object files should be shipped to customers. Or my be I didn't > notice configure option for intermediate build output :-) > > Solution with environment variable will decrease size of shipped > package, since only final binaries and support files will be shipped. > It may be more preferable solution when software installation script > is not trivial. Ok, I do see some value in that. I've created a ticket for it: https://trac.mpich.org/projects/mpich/ticket/1777 -- Pavan -- Pavan Balaji http://www.mcs.anl.gov/~balaji

1 0