diff --git a/s3tests/common.py b/s3tests/common/__init__.py
similarity index 81%
rename from s3tests/common.py
rename to s3tests/common/__init__.py
index 373d191..8eb4232 100644
--- a/s3tests/common.py
+++ b/s3tests/common/__init__.py
@@ -4,13 +4,24 @@ import itertools
 import os
 import random
 import string
+import sys
 import yaml
 
 s3 = bunch.Bunch()
 config = bunch.Bunch()
 prefix = ''
 
+# For those scripts that use a context, these are pretty univerally needed.
+context = bunch.Bunch(
+    bucket = None,
+
+    # Save stdout/stderr in case they get fudged with.
+    real_stdout = sys.stdout,
+    real_stderr = sys.stderr,
+)
+
 bucket_counter = itertools.count(1)
+key_counter = itertools.count(1)
 
 def choose_bucket_prefix(template, max_len=30):
     """
@@ -44,13 +55,19 @@ def nuke_prefixed_buckets():
                 print 'Cleaning bucket {bucket}'.format(bucket=bucket)
                 try:
                     bucket.set_canned_acl('private')
-                    for key in bucket.list():
-                        print 'Cleaning bucket {bucket} key {key}'.format(
-                            bucket=bucket,
-                            key=key,
-                            )
-                        key.set_canned_acl('private')
-                        key.delete()
+                    # TODO: deleted_cnt and the while loop is a work around for rgw
+                    # not sending the 
+                    deleted_cnt = 1
+                    while deleted_cnt:
+                        deleted_cnt = 0
+                        for key in bucket.list():
+                            print 'Cleaning bucket {bucket} key {key}'.format(
+                                bucket=bucket,
+                                key=key,
+                                )
+                            key.set_canned_acl('private')
+                            key.delete()
+                            deleted_cnt += 1
                     bucket.delete()
                 except boto.exception.S3ResponseError as e:
                     # TODO workaround for buggy rgw that fails to send
@@ -156,3 +173,12 @@ def get_new_bucket(connection=None):
 
 def teardown():
     nuke_prefixed_buckets()
+
+def fill_pools(*args):
+    for pool in args:
+        while not pool.full():
+            pool.spawn()
+
+def get_next_key(bucket=None):
+    return bucket.new_key("seqkey-{num}".format(num=next(key_counter)))
+
diff --git a/s3tests/common/greenlets.py b/s3tests/common/greenlets.py
new file mode 100644
index 0000000..e4e52bc
--- /dev/null
+++ b/s3tests/common/greenlets.py
@@ -0,0 +1,93 @@
+import bunch
+import collections
+import gevent
+import random
+import time
+import traceback
+
+from ..common import context, get_next_key
+from ..common.results import TransferGreenletResult
+from ..realistic import FileVerifier
+
+
+# Make sure context has somewhere to store what we need
+context.update(bunch.Bunch(
+    needs_first_read = collections.deque(),
+    all_keys = [],
+    files_iter = None,
+))
+
+
+class SafeTransferGreenlet(gevent.Greenlet):
+    def __init__(self, timeout=120):
+        gevent.Greenlet.__init__(self)
+        self.timeout = timeout
+        self.result = None
+
+    def _run(self):
+        result = self.result = TransferGreenletResult(self.type)
+        result.markStarted()
+
+        try:
+            with gevent.Timeout(self.timeout, False):
+                result.success = self._doit()
+        except gevent.GreenletExit:
+            return
+        except:
+            result.setError(show_traceback=True)
+
+        result.markFinished()
+
+
+class ReaderGreenlet(SafeTransferGreenlet):
+    type = 'reader'
+
+    def _doit(self):
+        if context.needs_first_read:
+            key = context.needs_first_read.popleft()
+        elif context.all_keys:
+            key = random.choice(context.all_keys)
+        else:
+            time.sleep(1)
+            return self.result.setError('No available keys to test with reader. Try again later.')
+
+        # Copynew the key object
+        key = key.bucket.new_key(key.name)
+        self.result.setKey(key)
+
+        fp = FileVerifier()
+
+        key.get_contents_to_file(fp)
+
+        self.result.request_finish = time.time()
+        self.result.request_start = fp.created_at
+        self.result.chunks = fp.chunks
+        self.result.size = fp.size
+
+        if not fp.valid():
+            return self.result.setError('Failed to validate key {name!s}'.format(name=key.name))
+
+        return True
+
+
+class WriterGreenlet(SafeTransferGreenlet):
+    type = 'writer'
+
+    def _doit(self):
+        key = get_next_key(context.bucket)
+        self.result.setKey(key)
+
+        fp = next(context.files_iter)
+        self.result.size = fp.size
+
+        key.set_contents_from_file(fp)
+
+        self.result.request_finish = time.time()
+        self.result.request_start = fp.start_time
+        self.result.chunks = fp.last_chunks
+
+        # And at the end, add to needs_first_read and shuffle
+        context.needs_first_read.append(key)
+        context.all_keys.append(key)
+
+        return True
diff --git a/s3tests/common/results.py b/s3tests/common/results.py
new file mode 100644
index 0000000..b9aa9ce
--- /dev/null
+++ b/s3tests/common/results.py
@@ -0,0 +1,112 @@
+import bunch
+import collections
+import gevent
+import sys
+import time
+import traceback
+import yaml
+
+from ..common import context
+
+context.update(bunch.Bunch(
+    result_queue = collections.deque(),
+))
+
+
+class TransferGreenletResult(object):
+    """ Generic container object. Weeeeeeeeeeeeeee *short* """
+    def __init__(self, type):
+        # About the Greenlet
+        self.type = type
+
+        # About the key
+        self.bucket = None
+        self.key = None
+        self.size = None
+
+        # About the job
+        self.success = False
+        self.error = None
+
+        self.start_time = None
+        self.finish_time = None
+
+        self.duration = None
+        self.latency = None
+
+        self.request_start = None
+        self.request_finish = None
+
+        self.chunks = None
+
+    def markStarted(self):
+        self.start_time = time.time()
+
+    def markFinished(self):
+        self.finish_time = time.time()
+        self.duration = self.finish_time - self.start_time
+        context.result_queue.append(self)
+
+    def setKey(self, key):
+        self.key = key.name
+        self.bucket = key.bucket.name
+
+    def setError(self, message='Unhandled Exception', show_traceback=False):
+        """ Sets an error state in the result, and returns False... example usage:
+
+        return self.result.setError('Something happened', traceback=True)
+        """
+        self.error = dict()
+        self.error['msg'] = message
+        if show_traceback:
+            self.error['traceback'] = traceback.format_exc()
+        return False
+
+    @classmethod
+    def repr_yaml(c, dumper, self):
+        data = dict()
+        for x in ('type', 'bucket', 'key', 'chunks'):
+            data[x] = self.__dict__[x]
+
+        # reader => r, writer => w
+        data['type'] = data['type'][0]#chunks
+
+        # the error key must be present ONLY on failure.
+        assert not (self.success and self.error)
+        if self.success:
+            assert self.error == None
+        else:
+            assert self.error != None
+            data['error'] = self.error
+
+        data['start'] = self.request_start
+        if self.request_finish:
+            data['duration'] = 1000000000 * (self.request_finish - self.request_start)
+
+        return dumper.represent_dict(data)
+
+# And a representer for dumping a TransferGreenletResult as a YAML dict()
+yaml.add_representer(TransferGreenletResult, TransferGreenletResult.repr_yaml)
+
+
+class ResultsLogger(gevent.Greenlet):
+    """ A quick little greenlet to always run and dump results. """
+    def __init__(self):
+        gevent.Greenlet.__init__(self)
+        self.outfile = context.real_stdout
+
+    def _run(self):
+        while True:
+            try:
+                self._doit()
+            except:
+                print "An exception was encountered while dumping the results... this shouldn't happen!"
+                traceback.print_exc()
+            time.sleep(0.1)
+
+    def _doit(self):
+        while context.result_queue:
+            result = context.result_queue.popleft()
+            yrep = yaml.dump(result)
+            self.outfile.write(yrep + "---\n")
+
diff --git a/s3tests/realistic.py b/s3tests/realistic.py
index e5d234b..0755516 100644
--- a/s3tests/realistic.py
+++ b/s3tests/realistic.py
@@ -1,18 +1,25 @@
+import bunch
 import hashlib
 import random
 import string
 import struct
+import time
 
 class RandomContentFile(object):
     def __init__(self, size, seed):
+        self.size = size
         self.seed = seed
         self.random = random.Random(self.seed)
-        self.offset = 0
-        self.buffer = ''
-        self.size = size
-        self.hash = hashlib.md5()
-        self.digest_size = self.hash.digest_size
-        self.digest = None
+
+        # Boto likes to seek once more after it's done reading, so we need to save the last chunks/seek value.
+        self.last_chunks = self.chunks = None
+        self.last_seek = self.start_time = None
+
+        # Let seek initialize the rest of it, rather than dup code
+        self.seek(0)
+
+    def _mark_chunk(self):
+        self.chunks.append([self.offset, (time.time() - self.last_seek) * 1000000000])
 
     def seek(self, offset):
         assert offset == 0
@@ -20,6 +27,17 @@ class RandomContentFile(object):
         self.offset = offset
         self.buffer = ''
 
+        self.hash = hashlib.md5()
+        self.digest_size = self.hash.digest_size
+        self.digest = None
+
+        # Save the last seek time as our start time, and the last chunks
+        self.start_time = self.last_seek
+        self.last_chunks = self.chunks
+        # Before emptying.
+        self.last_seek = time.time()
+        self.chunks = []
+
     def tell(self):
         return self.offset
 
@@ -58,6 +76,8 @@ class RandomContentFile(object):
             data = self.digest[:digest_count]
             r.append(data)
 
+        self._mark_chunk()
+
         return ''.join(r)
 
 class FileVerifier(object):
@@ -65,6 +85,11 @@ class FileVerifier(object):
         self.size = 0
         self.hash = hashlib.md5()
         self.buf = ''
+        self.created_at = time.time()
+        self.chunks = []
+
+    def _mark_chunk(self):
+        self.chunks.append([self.size, (time.time() - self.created_at) * 1000000000])
 
     def write(self, data):
         self.size += len(data)
@@ -72,6 +97,7 @@ class FileVerifier(object):
         digsz = -1*self.hash.digest_size
         new_data, self.buf = self.buf[0:digsz], self.buf[digsz:]
         self.hash.update(new_data)
+        self._mark_chunk()
 
     def valid(self):
         """
@@ -123,7 +149,54 @@ def names(mean, stddev, charset=None, seed=None):
     while True:
         while True:
             length = int(rand.normalvariate(mean, stddev))
-            if length >= 0:
+            if length > 0:
                 break
         name = ''.join(rand.choice(charset) for _ in xrange(length))
         yield name
+
+def files_varied(groups, unlimited=False):
+    """ Yields a weighted-random selection of file-like objects. """
+    # Quick data type sanity.
+    assert groups and isinstance(groups, (list, tuple))
+
+    total_num = 0
+    file_sets = []
+    rand = random.Random(time.time())
+
+    # Build the sets for our yield
+    for num, size, stddev in groups:
+        assert num and size
+
+        file_sets.append(bunch.Bunch(
+            num    = num,
+            size   = size,
+            stddev = stddev,
+            files  = files(size, stddev, time.time())
+        ))
+        total_num += num
+
+    while True:
+        if not total_num:
+            raise StopIteration
+
+        num = rand.randrange(total_num)
+
+        ok = 0
+        for file_set in file_sets:
+            if num > file_set.num:
+                num -= file_set.num
+                continue
+
+            if not unlimited:
+                total_num -= 1
+                file_set.num -= 1
+
+                # None left in this set!
+                if file_set.num == 0:
+                    file_sets.remove(file_set)
+
+            ok = 1
+            yield next(file_set.files)
+
+        if not ok:
+            raise RuntimeError, "Couldn't find a match."
diff --git a/s3tests/roundtrip.py b/s3tests/roundtrip.py
new file mode 100644
index 0000000..69b35a1
--- /dev/null
+++ b/s3tests/roundtrip.py
@@ -0,0 +1,99 @@
+import gevent.monkey
+gevent.monkey.patch_all()
+
+import bunch
+import collections
+import gevent
+import gevent.pool
+import itertools
+import random
+import realistic
+import time
+import traceback
+import sys
+
+import common
+from common import context, config
+from common.greenlets import ReaderGreenlet, WriterGreenlet
+from common.results import ResultsLogger
+
+# Set up the common context to use our information. Wee.
+context.update(bunch.Bunch(
+    # Set to False when it's time to exit main loop.
+    running = True,
+
+    # The pools our tasks run in.
+    greenlet_pools = bunch.Bunch(
+        writer=None,
+        reader=None,
+        ),
+
+    # The greenlet that keeps logs going.
+    results_logger = None,
+))
+
+
+def setup():
+    config_rt = config.roundtrip
+
+    context.bucket = common.get_new_bucket()
+    print "Using bucket: {name}".format(name=context.bucket.name)
+
+    context.greenlet_pools.reader = gevent.pool.Pool(config_rt.pool_sizes.reader, ReaderGreenlet)
+    context.greenlet_pools.writer = gevent.pool.Pool(config_rt.pool_sizes.writer, WriterGreenlet)
+
+    context.key_iter = itertools.count(1)
+    context.files_iter = realistic.files_varied(config_rt.create_objects)
+
+
+def _main():
+    def _stop_running():
+        """ Since we can't do assignment in a lambda, we have this little stub """
+        context.running = False
+
+    grace_period = config.roundtrip.grace_wait
+
+    print "Launching/Scheduling essential services..."
+    gevent.spawn_later(config.roundtrip.duration + grace_period, _stop_running)
+    context.results_logger = ResultsLogger.spawn()
+
+    print "Launching the pool of writers, and giving them {grace} seconds to get ahead of us!".format(grace=grace_period)
+    writers_start_time = time.time()
+    while time.time() - writers_start_time < grace_period:
+        common.fill_pools(context.greenlet_pools.writer)
+        time.sleep(0.1)
+
+    # Main work loop.
+    print "Starting main work loop..."
+    while context.running:
+        common.fill_pools(*context.greenlet_pools.values())
+        time.sleep(0.1)
+
+    print "We've hit duration. Time to stop!"
+    print "Waiting {grace} seconds for jobs to finish normally.".format(grace=grace_period)
+    time.sleep(grace_period)
+
+    print "Killing off any remaining jobs."
+    context.greenlet_pools.reader.kill()
+    context.greenlet_pools.writer.kill()
+
+    print "Waiting 10 seconds for them to finish dying off and collections to complete!"
+    time.sleep(10)
+
+    print "Killing essential services..."
+    context.results_logger.kill()
+
+    print "Done!"
+
+
+def main():
+    sys.stdout = sys.stderr # Original steam already saved by common
+    common.setup()
+    setup()
+
+    # Normal
+    try:
+        _main()
+    except:
+        traceback.print_exc()
+    common.teardown()
diff --git a/setup.py b/setup.py
index de01ab6..686fadd 100644
--- a/setup.py
+++ b/setup.py
@@ -23,6 +23,7 @@ setup(
         'console_scripts': [
             's3tests-generate-objects = s3tests.generate_objects:main',
             's3tests-test-readwrite = s3tests.rand_readwrite:main',
+            's3tests-test-roundtrip = s3tests.roundtrip:main',
             ],
         },