latching fixes, server no longer ignores --test, performance tuning

git-svn-id: svn+ssh://svn.corp.yahoo.com/yahoo/yrl/labs/pnuts/code/logstore@759 8dad8b1f-cf64-0410-95b6-bcf113ffbcfe

datapage.cpp

@@ -85,7 +85,7 @@ DataPage<TUPLE>::DataPage(int xid, pageid_t page_count, RegionAllocator *alloc)
 	first_page_(alloc_->alloc_extent(xid_, page_count_)),
 	write_offset_(0)
 {
-	DEBUG("Datapage page count: %lld pid = %lld\n", (long long int)page_count_, (long long int)first_page_);
+	printf("Datapage page count: %lld pid = %lld\n", (long long int)initial_page_count_, (long long int)first_page_);
     assert(page_count_ >= 1);
     initialize();
 }
@@ -241,7 +241,12 @@ template <class TUPLE>
 bool DataPage<TUPLE>::append(TUPLE const * dat)
 {
   // Don't append record to already-full datapage.  The record could push us over the page limit, but that's OK.
-        if(write_offset_ > (initial_page_count_ * PAGE_SIZE)) { return false; }
+  if(write_offset_ > (initial_page_count_ * PAGE_SIZE)) {
+    DEBUG("offset %lld closing datapage\n", write_offset_);
+    return false;
+  }
+
+  DEBUG("offset %lld continuing datapage\n", write_offset_);
 
 	byte * buf = dat->to_bytes(); // TODO could be more efficient; this does a malloc and memcpy.  The alternative couples us more strongly to datapage, but simplifies datapage.
 	len_t dat_len = dat->byte_length();

diskTreeComponent.cpp

@@ -805,12 +805,12 @@ diskTreeComponent::internalNodes::iterator::iterator(int xid, recordid root, con
     current.size = lsm_entry_rid.size;
 
     xid_ = xid;
-    t = 0; // must be zero so free() doesn't croak.
     justOnePage = (depth==0);
 
     DEBUG("diskTreeComponentIterator: index root %lld index page %lld data page %lld key %s\n", root.page, current.page, rec->ptr, key);
     DEBUG("entry = %s key = %s\n", (char*)(rec+1), (char*)key);
   }
+  t = 0; // must be zero so free() doesn't croak.
 }
 
 /**
@@ -859,7 +859,6 @@ int diskTreeComponent::internalNodes::iterator::next()
   } else {
     assert(!p);
     if(t != NULL) { free(t); t = NULL; }
-    t = 0;
     return 0;
   }
 }
@@ -869,8 +868,12 @@ void diskTreeComponent::internalNodes::iterator::close() {
   if(p) {
     unlock(p->rwlatch);
     releasePage(p);
+    p = NULL;
+  }
+  if(t) {
+    free(t);
+    t = NULL;
   }
-  if(t) free(t);
 }
 
 

logstore.h

@@ -36,7 +36,19 @@ class logtable {
 public:
 
   class iterator;
-    logtable(pageid_t internal_region_size = 1000, pageid_t datapage_region_size = 10000, pageid_t datapage_size = 40); // scans 160KB / 2 per lookup on average. at 100MB/s, this is 0.7 ms.  XXX pick datapage_size in principled way.
+
+  // We want datapages to be as small as possible, assuming they don't force an extra seek to traverse the bottom level of internal nodes.
+  // Internal b-tree mem requirements:
+  //  - Assume keys are small (compared to stasis pages) so we can ignore all but the bottom level of the tree.
+  //
+  //  |internal nodes| ~= (|key| * |tree|) / (datapage_size * |stasis PAGE_SIZE|)
+  //
+  // Plugging in the numbers today:
+  //
+  //  6GB ~= 100B * 500 GB / (datapage_size * 4KB)
+  //  (100B * 500GB) / (6GB * 4KB) = 2.035
+  logtable(pageid_t internal_region_size = 1000, pageid_t datapage_region_size = 10000, pageid_t datapage_size = 2);
+
     ~logtable();
 
     //user access functions
@@ -259,9 +271,11 @@ public:
       }
 
       ~iterator() {
+        writelock(ltable->header_lock,0);
         ltable->forgetIterator(this);
         invalidate();
         if(last_returned) TUPLE::freetuple(last_returned);
+        unlock(ltable->header_lock);
       }
   private:
       TUPLE * getnextHelper() {
@@ -292,6 +306,7 @@ public:
       }
 
       void invalidate() {
+        assert(!trywritelock(ltable->header_lock,0));
         if(valid) {
           delete merge_it_;
           merge_it_ = NULL;

merger.cpp

@@ -275,6 +275,7 @@ void* memMergeThread(void*arg)
     while(true) // 1
     {
         merge_stats_t stats;
+        memset((void*)&stats, 0, sizeof(stats));
         stats.merge_level = 1;
         stats.merge_count = merge_count;
         gettimeofday(&stats.sleep,0);
@@ -434,6 +435,7 @@ void *diskMergeThread(void*arg)
     while(true)
     {
         merge_stats_t stats;
+        memset((void*)&stats, 0, sizeof(stats));
         stats.merge_level = 2;
         stats.merge_count = merge_count;
         gettimeofday(&stats.sleep,0);

server.cpp

@@ -77,9 +77,9 @@ int main(int argc, char *argv[])
     int lindex = mscheduler->addlogtable(&ltable);
     ltable.setMergeData(mscheduler->getMergeData(lindex));
 
-    int64_t c0_size = 1024 * 1024 * 1024 * 1;
+    int64_t c0_size = 1024 * 1024 * 512 * 1;
 
-    if(argc == 2 && !strcmp(argv[0], "--test")) {
+    if(argc == 2 && !strcmp(argv[1], "--test")) {
 
       c0_size = 1024 * 1024 * 10;
       printf("warning: running w/ tiny c0 for testing"); // XXX build a separate test server and deployment server?
@@ -89,7 +89,7 @@ int main(int argc, char *argv[])
 
     unlock(ltable.header_lock);
 
-    lserver = new logserver(10, 32432);
+    lserver = new logserver(100, 32432);
 
     lserver->startserver(&ltable);