If you are using multiple connections to the Isilon,

are the connections evenly distributed across the nodes?

(SmartConnect can't help much if one client starts

numerous connections based on a single DNS query.

If there is one query per connection,

but all within a few seconds, round-robin

need to be used in SmartConnect rather than

load-based policies.)

Next, how exactly are the writes accomplished from the Netazza

(as compared to the better performing clients)?

Use isi statistics client (--long)

and identify the operations from the client

in question (RemoteName/RemoteAddr).

Check:

- In (and Out): throughput per "connection" (multiple connections between

    the same IPs (LocalName--RemoteName) a coalesced into a single line though).

- Class: read vs write vs namespace operations: streaming write loads

     should mostly show large 'write' ops and few if any non-write ops.

- Ops:  oprate/sec, very high (with small request size, InAvg)? very low (with large TimeAvg)?

- TimeAvg: time spent within the Isilon per request (in usec):

       10000usec=10msec would be considered  high.

- (with --long:) InAvg (and OutAvg): average request sizes, e.g. 4K would be too small for streaming...

In case you see mainly large writes (as considered good), compute

TimeAvg-in-seconds times Ops-per-second:

- TimeAvg/1000000 x Ops = Utilization rate per "connection"

    if substantially below 1, then latencies outside the Isilon occur, like

    network latencies or client-side delays/queuing problems.

(Edit 9/13: Utilization rate: here not the storage utilization w.r.t. protection

overhead, but utilization of a processing unit, Little's law - Wikipedia, the free encyclopedia)

Short story: The Netazza should show a pattern clearly different

from the ok clients. Once identified, appropriate

means can be investigated.

-- Peter