diff --git a/Feersum.xs b/Feersum.xs
index e456d82..952cac2 100644
--- a/Feersum.xs
+++ b/Feersum.xs
@@ -7,13 +7,10 @@
 #include <ctype.h>
 #include <netinet/ip.h>
 #include <netinet/tcp.h>
+#include <sys/uio.h>
 
 #include "ppport.h"
 
-#include "picohttpparser-git/picohttpparser.c"
-
-#include "rinq.c"
-
 #ifdef __GNUC__
 # define likely(x)   __builtin_expect(!!(x), 1)
 # define unlikely(x) __builtin_expect(!!(x), 0)
@@ -101,13 +98,21 @@
 #define trace3(...)
 #endif
 
-#include <sys/uio.h>
-#define IOMATRIX_SIZE 64
+#include "picohttpparser-git/picohttpparser.c"
+#include "rinq.c"
+
+// Check FEERSUM_IOMATRIX_SIZE against what's actually usable on this
+// platform.  See $iomatrix_size in Makefile.PL for the default.
+#if FEERSUM_IOMATRIX_SIZE > IOV_MAX
+# undef FEERSUM_IOMATRIX_SIZE
+# define FEERSUM_IOMATRIX_SIZE IOV_MAX
+#endif
+
 struct iomatrix {
     unsigned offset;
     unsigned count;
-    struct iovec iov[IOMATRIX_SIZE];
-    SV *sv[IOMATRIX_SIZE];
+    struct iovec iov[FEERSUM_IOMATRIX_SIZE];
+    SV *sv[FEERSUM_IOMATRIX_SIZE];
 };
 
 struct feer_req {
@@ -2613,4 +2618,15 @@ BOOT:
 
         Zero(&psgix_io_vtbl, 1, MGVTBL);
         psgix_io_vtbl.svt_get = psgix_io_svt_get;
+
+        trace3("Feersum booted, iomatrix %lu "
+                "(IOV_MAX=%u, FEERSUM_IOMATRIX_SIZE=%u), "
+            "feer_req %lu, "
+            "feer_conn %lu\n",
+            (long unsigned int)sizeof(struct iomatrix),
+                (unsigned int)IOV_MAX,
+                (unsigned int)FEERSUM_IOMATRIX_SIZE,
+            (long unsigned int)sizeof(struct feer_req),
+            (long unsigned int)sizeof(struct feer_conn)
+        );
     }
diff --git a/Makefile.PL b/Makefile.PL
index 7b147bd..e8efe02 100644
--- a/Makefile.PL
+++ b/Makefile.PL
@@ -58,6 +58,17 @@ if ($convert_deps) {
 use Config;
 my $steal = ($] >= 5.012 && !defined($Config{useithreads}));
 
+# Control the size of the main write-buffer structure in feersum.  Making this
+# value lower will use slightly less memory per connection at the cost of
+# speed and vice-versa.  It's most noticeable when you're app is buffering a
+# lot of sparce writes.  The default of 64 generally keeps usage under 4k per
+# connection on full 64-bit when you take into account the other connection
+# and request structures.
+# NOTE: If your OS defines IOV_MAX or UIO_MAXIOV then this value will be
+# forced to be at most that size with a #define.  Solaris defines IOV_MAX to
+# be 16.  Linux and OSX generally have IOV_MAX as 1024.
+my $iomatrix_size = 64;
+
 WriteMakefile(ev_args(
     NAME          => 'Feersum',
     VERSION_FROM  => 'lib/Feersum.pm',
@@ -92,7 +103,8 @@ WriteMakefile(ev_args(
     },
     LIBS   => [''],
     EXE_FILES => ['bin/feersum'],
-    DEFINE => ($steal ? '-DFEERSUM_STEAL' : ''),
+    DEFINE => "-DFEERSUM_IOMATRIX_SIZE=$iomatrix_size ".
+        ($steal ? '-DFEERSUM_STEAL' : ''),
     INC    => '-I.',
     dynamic_lib => {OTHERLDFLAGS => $otherldflags},
 ));