Quoth [email protected]:

> Allow me to disagree.  Arenas are, mostly, append-only.  Index is
write-random.  For each clump about 40 bytes are changed at random
locations.  Therefore, you are are safe filling up the arena disc
entirely,  but the index disc will be weared up, for nothing.  Most
ssd vendors recommend using a very low percentage of nominal capacity
for mit use cases.


> for each clump, about 40 bytes are changed

Average clump size is on the order of 10KiB, or ~20 sectors.  Looking
at orders of magnitude, we're talking about roughly one (random)
sector written to the index for every 20 (sequential) sectors written
to the arena.

Moreover, arena writes ALSO have random writes!  Each write to the
arena, in addition to the sequential clump, seeks to the end of the
arena to add the index entry so that, if the index dies, it can be
rebuilt without scanning the actual data area.  There should be one of
those for each write to the index.

Yes, the index write is _more_ random, but there's also _so few of
them_ that it shouldn't matter.

keep in mind as well that the index is intended to be sized such that
it never hits 100% of usage.

If you have actual data to show that this is a problem, I'm happy to
take a look, but it seems to be speculative and based on ideas of how
SSDs _should_ work, not based on any actual SSD that you can point to
and say "look, its worn out in less than a year from being used as an
index"?



------------------------------------------
9fans: 9fans
Permalink: 
https://9fans.topicbox.com/groups/9fans/T4d8ba4236feb5d92-Mafd2cbacdf7b4c1799ce1c77
Delivery options: https://9fans.topicbox.com/groups/9fans/subscription

Reply via email to